GitHub really needs something finer-grain then just pretending the repo never existed during these incidents. [1]
The bad package version has also just disappeared from crates.io [2] with no indication its been yanked. There's no security advisory there either [3] "No advisories found for this crate."
I feel crates.io was unprepared for a security incident like this since they're managing the response [4]
> The bad package version has also just disappeared from crates.io with no indication its been yanked.
So, yanked crate versions do have an indication on crates.io. (Here's an example: [1]) The Rust blog post uses the word "deleted", and I'm guessing a bit here, but I think they mean that literally, and that the version here is deleted, not yanked. And I think that would be more appropriate: a yanked crate is still downloadable by cargo, if your lockfile is locked to it already; yanking only prevents lockfiles from newly automatically acquiring a lock on that version. That wouldn't be desirable in the case of a compromised crate: you don't want a download occurring at all. The tradeoff of "break those with locks on the crate" tips to being worth it.
That said, I agree with you, though: I think this state (if it is "deleted" and not yanked) should be plainly indicated on the crate's versions page. (Even better would be if it came with a link to, e.g., the blog post or a RUSTSEC so that you could find out why.) (& I think perhaps the docs for yank should point out whether or not it is appropriate in the "compromised crate" scenario, and if not, what to do instead.)
> That wouldn't be desirable in the case of a compromised crate: you don't want a download occurring at all.
Sure you do; you just don’t want the crate to be made available to people trying to make use of the crate as regular downstream consumers. You still want the crate available for download for analysis. You especially want to be able to deterministically reproduce the vulnerable version of your software that you already built.
I would posit that most package ecosystems should treat “contains an exploit” as its own package state; mostly ignore them; but still be aware of their existence.
In other words, if there’s any other non-exploited option under your version constraints, the dep should resolve to the nearest non-exploited version, even if older. But if it’s the only version, or if you have pinned the exploited version, then the packager should see it, but normally refuse to interact with it—i.e. refuse to lock to it if it’s the only version; or refuse to fetch it if it’s the locked version.
I say “normally” because you should be able to bypass this with a flag / env-var to the packager that basically means “I’m building this for forensic analysis, not for running.”
I agree but think the problem is slightly more general. Consider that a package could be redacted not just due to exploit but also (among other things) for legal reasons, and that such reasons could be region specific. So a redacted package might or might not be available to manually download and some other metadata is desirable so that "audit" can notify you that you were compromised.
A package redacted for legal reasons should still be able to be built too! Presumably one or the other side of the relevant suit will need the package as evidence!
I would suggest that such packages should simply require authenticated access + manual “write an email to ask to have your account be put on a whitelist”-driven authorization. Such a package would still be there (in its redacted state); but attempts to download it would be a 401 (if unauthenticated) or 403 (if not on the whitelist) error response, which the packaging tooling would need to know how to handle.
(In any case, this “redacted” package state and my “exploited” package state could be two variants of the same logic, both following the same approach to locking/resolution, both requiring the flag to indicate that you understand what you’re doing and aren’t just going to run the resulting program. The only difference is that once you pass that flag, an “exploited” package would then just work, for anyone; while a “redacted” package would still have one more gate standing (server-side AAA) in the way.
I think "deleted" is still always going to have to be a package state, realistically. A package could contain spam, abuse, copyright infringement and/or illegal content.
I don't mind malware being lumped in to that - the place to look up code for historical or analysis reasons ought to be version control, not crates.io
What about the place to look up packaged binary releases for historical or analysis reasons, esp. when the packaging system either builds such releases itself, or where it allows the submission of arbitrary build artifacts (where the build process that produced said artifact can be compromised in ways that have nothing to do with the source repo)?
> the place to look up code for historical or analysis reasons ought to be version control, not crates.io
The build infrastructure should make security incidents easier to respond to, not harder. I shouldn't be sent on a wild goose chase at a time when I'm potentially already dealing with a major incident. The tools need to "just work".
[3] is no longer true. They're definitely not unprepared for an incident like this. It's not the first time they've done it and they published an update to their process in Feb:
> We shouldn't need to resort to pasting `find` commands [1] into the shell from blog posts to tell if we're compromised.
I know you recommended `cargo audit`, but that is not nearly as universally installed as `find`. And as far as I know, cargo-audit has to be run within a project directory? Can it be run outside of the context of a project and scan the cargo registry cache?
I personally appreciated having the `find` command:
* It very clearly indicates where to look (my cargo registry cache)
* It very clearly indicates what files to look for (a list of wildcards)
* It's something that I can easily review and then copy/paste into my terminal
* I can very easily adapt it to my particular environment (perhaps into a `fd` invocation if I'm on Windows, or perhaps adapt it to scan all home directories on my system), or feed these file names into some other vulnerability scanner
I believe what's being suggested is that `cargo audit` should be a part of cargo rather than an add on.
You'd obviously already have cargo installed, which means you should be able to run a command against cargo to see if you are currently exposed to any security problems. Even if it's not the current "audit" add on.
As someone using bazel at work, the find script was much more amenable to me trying to audit our usage - we don't use cargo, so no cargo cache or cargo audit, but I can still figure out whether a particular package was downloaded.
If you're using a different tool then it falls to that tool to address the same problem. The point is that I shouldn't have to already know that a compromise has occurred and then go looking for it manually. The tooling should automatically start shouting at me very loudly that something happened with one of the packages I'm using.
They did that because a yanked package can still be used. And in this case there weren't any downloads, so it is unlikely the malicious package was actually used.
Call it how you want, the point is that it should not disappear like this. Maybe yanked has the meaning you said, but in Ruby yanked has the meaning OP said and that is what we want.
How to call it then : Yanked hard vs yanked soft ?
The version page [1] should show that for 86 minutes there was a version 0.3.10, it was malicious and was deleted with a link to the advisory. I get this takes time so even a generic "deleted" entry until they have time to link in the advisory would also be fine so we know something is happening.
Presumably even "deleted" crate versions still have some metadata left behind in the backend so this should be surfaced.
> GitHub really needs something finer-grain then just pretending the repo never existed during these incidents.
Google should read this too. They simply remove Android apps and Chrome extensions without a single word. No page explaining why they removed it, if i was at risk.
I think we should be taking a more “batteries included” approach to language and library design. The entire reason we’re in this mess is because we’ve decided it’s ok or maybe even preferable if stdlibs are rail thin, rendering base languages near-unusable.
I can very easily build a highly functional, pleasant to use Apple platform app with 5 or fewer top level dependencies. In many cases, I reach for between 0-2 total.
There’s no reason why this can’t be replicated elsewhere. The key is to make the programming language reasonably robust with at least 80% of common non-UI dev needs built in and put the remaining 20% and UI bits into a small family of well-supported, community-embraced, preferably first party libraries.
That would make it unnecessary to pull in foreign dependencies in the overwhelming majority of projects. What few do get pulled in becomes lightweight, easily verifiable syntactic sugar or libraries with purposes too niche to be worth targeting.
Of course this approach can go wrong too. You could easily end up with a monster like Boost, but that comes down to project administration keeping creep under control and proper modular design.
This is spot on. I am extremely uncomfortable with the large number of transitive dependencies that end up in pretty much any non-trivial rust application. No amount of memory safety will save us if a tiny, ubiquitous library that nobody scrutinizes because it’s nested eight layers down the dependency graph gets compromised.
I think Go apps tend to have better dependency hygiene because the language has a better standard library, which results in better culture around dependencies.
I also think it’s frustrating that both cargo and npm totally ignore the decades of prior art from Linux distributions that have figured out good ways to improve dependency management. We’d be in better shape if there was a community curated subset of known-good dependencies that are release-managed together that the broader ecosystem could build on, sort of like Ubuntu having “main” and “universe”.
I learned programming from C# which has an excellent standard library. Just recently started using Go which seems to have a good standard library. Been using python for years, can't say I ever had a complaint about the standard library.
Rust meanwhile seems to be following some "no standard library" philosophy.
I feel like there's an opportunity out there to become like an amoeba: fund and build a third party Rust standard library, absorb absolutely everyone who is so grateful to have a library, and then get so big absorb Rust itself.
I think Go ends up with fewer dependencies mostly because there is more friction in finding them. With rust (and npm) you just "cargo search xxx", then "cargo add xxx". Go makes you web search and poke around github looking for something. It's not onerous or anything, but just that extra step slows things down just a bit. C projects tend to have the least because it's even more annoying to add them and you have to create build commands that deal with slightly different distributions (and learn makefile+pkg-config or cmake, or autoconf).
Certainly having a larger stdlib helps Go projects keep their deps down, but I don't think it's the real reason.
Perhaps, but I think it depends on where the Go devs are coming from. In my experience the lack of proper dependency management in older versions of Go didn't really lessen dependencies, it just made teams have to deal with annoying $GOSRC issues. But then I worked at a place where the devs used PHP (w/composer) and node before Go was introduced.
Personally I came from a C background so I tended to use deps more sparingly.
In the end it's a bit of a balancing act—lots of dependencies is a larger opportunity for these kind of supply chain attacks to affect you. Copying stuff into your repo protects you from that, but it also makes it way more likely that you'll miss security fixes, unless you're actively looking for them. Re-implementing what you need is also viable but it slows down the dev process.
I agree. Java has the same ability to quickly add deps, and in some case they can be sprawling, but it's still perfectly possible to develop complex apps without a lot of deps coming in.
> Certainly having a larger stdlib helps Go projects keep their deps down, but I don't think it's the real reason
I disagree: having a big stdlib that encompasses a huge chunk of common functionality (logging, a plethora of network server/client protocols, etc) means that for basic needs, the question of 3rd party libs never arise, whereas for other language, the question is not if you need a dep, but which one. Making gorilla easier to find won't undo the cultural reluctance of adding a 3rd party dep when using the (very pragmatic, IMO) stdlib server is adequate to the task.
>There’s no reason why this can’t be replicated elsewhere. The key is to make the programming language reasonably robust with at least 80% of common non-UI dev needs built in and put the remaining 20% and UI bits into a small family of well-supported, community-embraced, preferably first party libraries.
The big reason is money & time. Maybe less so today if you are happy with an AI generated stdlib, but I don't think it's fair to compare Rust (built by Mozilla, who's primary product is an open source web browser) to Swift/AppKit (built by Apple, 4.6T marketcap at the time of writing).
When Go was released it had an amazing stdlib, but that was because Google was funding it.
There's probably to truth to this, but I would weight the design intention behind the language just as or more heavily than available resources.
Swift was intended from the start to not only replace Objective-C and all of its use cases, but also take on new use cases and bring a number of features from other languages that were newer and had different dominant paradigms.
Certainly having a juggernaut like Apple behind it has been instrumental in its success in achieving those goals. That said, a community-lead project with similar aims could probably achieve those aims as well, given enough time.
I think it's just rare for enthusiast-led projects to set out to do such things. My theory on why things tend to go that way is that the types of people to start programming language projects tend to heavily lean nuts-and-bolts and theory with a purist/idealist sort of mentality that's more concerned ideological purity than practical usability.
So to sum this all up, a community-run practical, multi-purpose, batteries-included language likely needs to at least partially designed and helmed by product engineer types so it doesn't end up anemic and stuck in an ideological rut.
Python is batteries included, but all the batteries have corroded.
Look at C++'s long in the tooth STL.
You don't want to marry a language to fast aging libraries you have to support for eternity. Better libraries always naturally emerge.
Rust's decision here is fine.
The only thing I'd like to see is the ability to programmatically limit transitive dependency count or depth in Cargo. I'd also like crates to specify whether they limit their own deps and whether they have panick-y behavior or not.
I think maybe the right answer is to have a standardized, curated group of libraries pegged at some sort of LTS release that only backports security fixes. However, someone would need to pay for creating and maintaining this -- and then you wonder where the money would come from?
That or potentially fast tracking merging popular libraries into the standard library. Or at least concepts from popular libraries. Basically all the "most downloaded" crates on the Crates.io front page should be candidates for merging into the standard library.
Not everything should be eligible for this treatment. Certainly not an HTTP library or something without extremely widespread applicability.
Nearly all of my projects used once_cell, so it's good to see that pulled in. I wouldn't want to see anyhow, thiserror, anything opinionated, or anything domain specific in the std. That's a weight you have to bear forever.
Remember how long Python 2 -> 3 took, and look at the ancient and awful stuff in Python 3's standard library. Rust is not the right language for this.
I’m a big fan of how C# handles this. The standard library is versioned like any other library. Your program is explicit about the major version of the standard library you’re pulling in. And all major versions of the standard library continue to work.
For rust, this would mean something like adding std = 1.0.6 to your cargo.toml. The advantage is it means the standard library can deprecate and replace features. Upgrading std is an explicit step by the developer. Annoying, but in the age of LLMs it should be pretty easy.
The one big question I can’t answer is what happens if your dependencies use a different version of std? Are std v1 Strings compatible with std v2 strings? Oof.
Can we pick and choose the good without the bad? If we agree stability is a net-negative we could just say there will be a list of blessed trustworthy libraries with extra scrutiny and bureaucracy (make no mistake this is completely necessary to what people are asking for) to make sure they don't go rogue. But without a guarantee they will stick around and get updates forever.
Rust has a sizable and well featured standard library at this point. I think it would be absurd to claim that the base language is "near-unusable" if you are using it for system programming, which is its intended use case.
Huge standard libraries make sense for Java, Go, Apple platform etc. because they are developed by giant enterprises that can manage the overhead. Thinner modular systems are a more natural fit for open source development.
True. And the rust standard library is pretty good. But I really wish we had a good, fast, small futures executor in std. And accompanying async variants of File and Socket and so on.
Async rust is a jungle of weird compatibility questions. People treat async runtimes like sports teams. (I know I do). I wish it were more like Nodejs where async is just built in.
It’s being used for a wide variety of other purposes however, and so perhaps it’s time to adapt to that reality.
While overhead is a real concern, I think it’s often blown out of proportion. Once a language achieves a certain baseline of stability and isn’t in constant flux and the standard library matures, changes become infrequent and maintenance load is low. The work is heavily front-loaded.
I don't see the problem with boost? Isn't that a perfect example of your well supported, community embraced option? I certainly feel much safer pulling something in from boost via the official debian repos than I do pulling a random package with npm or cargo or etc.
Primarily I think the underlying concern has to do with the pathway for authoring code. When you have contributors whose submissions are gated with a rigid third party process and where that third party is the one responsible for curating the code (as opposed to the author also being the curator as well as the publisher) then you have the possibility to catch a lot of wrongdoing before it succeeds.
There are many issues that I've seen taken up with Boost, but my personal peeve is how it can make building projects that incorporate it a pain, both because of its sheer size but also because it can sometimes be difficult to appease with its own dependencies.
> I can very easily build a highly functional, pleasant to use Apple platform app with 5 or fewer top level dependencies. In many cases, I reach for between 0-2 total.
> There’s no reason why this can’t be replicated elsewhere.
It is replicated elsewhere and is not limited to Apple; I can do the same using a tech stack of perhaps 2x techs and very limited (manually added) deps. I'm thinking Lazarus, which comes with most things I'd need for most apps.
> I can very easily build a highly functional, pleasant to use Apple platform app with 5 or fewer top level dependencies. In many cases, I reach for between 0-2 total.
> There’s no reason why this can’t be replicated elsewhere. The key is to make the programming language reasonably robust with at least 80% of common non-UI dev needs built in and put the remaining 20% and UI bits into a small family of well-supported, community-embraced, preferably first party libraries.
I wholeheartedly agree. Most standard libraries out there (when I say most I easily mean of 99% languages that I know) have pitifully tiny implementations. They at most provide for string parsing, io, a bare networking layer (that's at best a wrapper around the POSIX syscalls), threading and that's it. If you want to do anything more complex, you either have to roll it yourself or use an assortment of the "blessed libs" for any given language, that may or may not work properly, and may or may not even work well with each other. I got pissed off by this at some point so I decided to go ahead and roll my own standard lib that would actually have everything I need to.. actually develop apps. Now I know most people probably aren't going to be using it (mainly because its functional API isn't what most C++ devs want) but for me it's actually been brilliant.
This was actually the main thing that put me off of Rust. I get the argument for a small std lib. I just also don’t agree it’s worth it. Go seems to handle having a batteries included standard lib just fine.
They do not want a big stdlib. The downsides are real (the stdlib cannot make breaking changes), and there are no upsides (except, maybe, for faster compilation, since std comes precompiled).
They want more official crates (e.g. `regex` and `libc` are official crates, maintained by the Rust project). And the Rust project does not oppose to that, it just doesn't have the funding.
> there are no upsides (except, maybe, for faster compilation
There are other good reasons to have something in std. Another big one is providing types for cross-crate compatibility. Like if I want to pass a String from one crate to another, I’m glad that string is defined in std so there’s an obvious type we can both use in our APIs. C - for example - does not have this luxury.
It would be nice if we had a standard way to declare a type as serialisable. Right now lots of crates have a serde feature to do this with serde - but they need to put this behind a feature flag (for obvious reasons) and this doesn’t work with other serialisation libs.
The other big one is futures. There’s still no futures executor in std. Async crates have to decide if they want to tie themselves to a single executor (like Tokio) or be compatible. There is no async stream api in std. No async file api. And so on. It’s a compatibility nightmare.
Who said a standard library can't make breaking changes? That used to be a norm some time ago.
There are upsides, your program is smaller, better security because a random kid can't pwn your deps, quality and interoperability. What's not to love?
>Most mainstream languages guarantee stability, certainly Rust.
That's their mistake, right there... When you're upgrading software, you presumably want a better version. You can't have something better without changing it. It's a logical contradiction.
>It doesn't matter if the code is in std or a crate.
It very much does because you don't need 95% of the random stuff in the crates. Even for something like rand, you need an xorshift and that's roughly it. 25 lines of code, sorted. You can even write it as a copypaste "dependency" without much fuss in practically any language. That, versus importing a whole rand library with lots of different algorithms, deps on crypto, tests, OS random-based seeding, customisability, etc.
>you just need more official crates
I guess that's one solution but it would probably just be a better idea to split std in two SLAs, one is guaranteed for core stuff i.e. the status quo, one is YMMV.
> Who said a standard library can't make breaking changes?
I don't want my language creating work for me. "Batteries included" is a feature of higher level languages. Rust is going to live for a very long time, probably beyond all of our lifetimes, and it needs to endure fads without picking up baggage and liabilities.
Rust could have a mandate where certain crates are "blessed" as recommendations. This could come with additional review, oversight, support, documentation, etc.
The Cargo.toml spec should have the ability to limit direct and transitive dependency count and depth. (As well as other things, like "nopanic" annotations/guarantees.)
The Crates.io repo should grow namespaces so we don't typo squat popular packages. It's much harder to miss organization names changing than package names changing.
Yeah, but I will take a not great library that works everywhere the compiler does, than be at the whims of which platforms are supported by 3rd party libraries.
I can use most of the clunky Python, Java, .NET and if it must be, Go, standard libraries, than hunting down for dependencies with platform tier support and such.
Naturally there is a balance, still I think Java, .NET, Python, Go, Smalltalk, manage quite well, while having subsets to slim down when needed for specific deployment scenarios like embedded devices.
Platform support is a non-issue for every library being discussed. This is an array construction macro, it has nothing to do with what operating system you're running on.
which, as painful as it may be, is ok. Better to have a safe functional stdlib library than a exposed external crate (which then asks the question; what's the replacement...)
We also have the classic example of PHP with numerous not safe stdlib ways to use mysql.
To me the answer is still to vendor your dependencies and don't be on the bleeding edge of updates unless you're willing to invest the time into validating them.
What is your critique of their approach? Is it not the case that `go get` is the only one which doesn't even provide a way for the person downloading to run the downloaded code until it is actually executed by the consuming codebase? That seems pretty sound by comparison.
I guess you're saying, "they had to" meaning they should've seen the need and provided it? I'll say this to that (imagined) take; plenty of useful software was made without it, and they got the job done when they knew the absolute most about what a good solution would need. I totally understand being annoyed at the fact that this is the story of every evolution in Go, but I genuinely think they're picking good implementations when they decide on them.
It’s half arsed, brittle and far from user friendly.
> Is it not the case that `go get` is the only one which doesn't even provide a way for the person downloading to run the downloaded code until it is actually executed by the consuming codebase?
If you’ve added the package to your imported then odds are your next step is going to build it. Thus negating any benefit.
I think the real issue is malicious packages entering package ecosystems. Whether your package manage executes on downloads or not is moot because you’ve still got untrusted code sat in your project imports, just waiting to be accidentally executed.
> I guess you're saying, "they had to" meaning they should've seen the need and provided it? I'll say this to that (imagined) take; plenty of useful software was made without it, and they got the job done when they knew the absolute most about what a good solution would need. I totally understand being annoyed at the fact that this is the story of every evolution in Go.
I think you’re being too charitable here. I think Russ Cox just didn’t want a package manager because C doesn’t have one. But ended up relenting after everyone nagged the Go team for years afterwards.
And really what they built was the bare minimum to manage package version pinning and updates. But it has none of the visibility that central package repositories have. So how do you know if a Go package has been compromised when the only source of truth is the compromised origin?
Ultimately I think devs need to think about their dependencies and decide which ones get pinned and a serious review before pulling. If the thing has got binaries, it gets a serious review. If it does anything with cryptography, it gets a serious review... etc.
I don't think automatic updating is a good idea at all. It's just the honor-system, and trust is a security flaw.
Thanks for the thorough reply. I personally haven't experienced brittleness/unfriendliness, but I have only written around 10k lines of Go. Not exactly a power user, but I like to think I understand it.
> If you’ve added the package to your imported then odds are your next step is going to build it. Thus negating any benefit.
I actually think this is the benefit! I don't need to go to a website to see what changed, I can just have a look at the code. The best dependencies have a changelog. Ideally I can look at a diff.
> I think you’re being too charitable here. I think Russ Cox just didn’t want a package manager because C doesn’t have one. But ended up relenting after everyone nagged the Go team for years afterwards.
That's entirely possible.
> So how do you know if a Go package has been compromised when the only source of truth is the compromised origin?
Fair point, and probably worse in this future we're in now that NIST is drowning in CVEs and has turned away from some share of them.
NPM came along in 2010, Javascript was huge before that.
It was only when people wanted a common approach to shipping both in browser and "native" that this glitch happened. I believe a standard library and "batteries" would have abated it entirely or resulted in a slightly less-bad situation. I do think Cargo is a slightly less-bad situation in many ways, and that it can be done better.
Sure. Maybe threshold is bit higher than moderately. But unless your language tries to sabotage itself by making code artifacts uncomposable (a la C/C++ where best way to compose libraries is through shell commands) some package manager will be inevitable.
Batteries also don't help if dependencies don't replace them. Arrayref functionality has been part of Rust std lib for a while now.
I don't like the "bring your own" approach because unless one happens to be an absolute tour de force 10x engineer unstoppable god of a programmer (which most of us, myself included, are not), whatever you build is never going to be as well-rounded, fleshed out, and complete as something built by a larger organization, especially when it comes to UI libraries (which are monstrous projects if done right, e.g. meeting accessibility requirements).
I'm happy to contribute to a larger effort but anything I can build on my own is going to be a thin, flimsy happy meal toy compared to something with the backing of a company or well organized FOSS project.
I agree that building your own is hard and something that can be done by exceptional engineers, but software is famously a winner-take-all industry, even if the optimal strategy for lower percentile programmers and median programmers to import generic modules, as long as building your own results in the best product (which it does, whether importing frontend or backend modules, if you hand code something, it's going to be the optimal strategy for building a winner product. Even programmers that are median or below average probably have an incentive to aim for being exceptional, because the best value of median and below average programmers is not building median or low value software, but having a shot at building exceptional software.
Also the challenge of building a generic module is much larger than building your own. A generic module needs to have flexibility for many different options and integrations, when you build your own you build just what you need and tightly integrate it with the product
> Also the challenge of building a generic module is much larger than building your own. A generic module needs to have flexibility for many different options and integrations, when you build your own you build just what you need and tightly integrate it with the product
True but I believe overestimated. Usually rather than building what's needed, what gets built is what is thought to be needed (often a substantially smaller subset), and then over time you end up building a markedly poor version of a generic module.
with LLMs it's becoming more common to just vibe up anything you need that might be missing. even if there is an available package you can pull in. the most secure option as well
Sandboxing is a mitigation but a very limited one. The library itself can contain a malicious payload -- rust code most often ends up as native code executed on the host.
I think we may need to enter a world where there are fewer, heavily audited, libraries and dependency depth is limited. Allowing unvetted dependencies to be installed by default is not a good way forward. App stores have a similar problem and I think a lot of lessons have been learned there that can be applied.
That world can’t exist unfortunately, because the minute someone wants a feature that isn’t in your core libraries, you just create a new library and we’re back to now.
This is exactly what PMG is designed for ie. install/build time process level sandboxing. It currently doesn't support cargo, but I believe the challenges are same.
Here is my learning building PMG:
Sandboxing is good when the workload is predictable, and the goal of sandbox is to guard against exploitation of vulnerabilities, like sandbox protecting chrome tabs (renderers). But unfortunately build scripts are not predictable, at least not in npm/pypi world and I have seen build scripts doing weirdest of the things which is no different from malware. When popular packages do weird things, build breaks and users end up turning off the sandbox. This is a perpetual problem to deal with while building sandbox (or any least privilege solution) to protect unbounded workloads.
build.rs by design can run absolutely anything. There tons of build.rs scripts that invoke a whole-ass C compiler toolchain to build and link C dependencies...
It isn't so much a question of sandboxing build.rs, as fundamentally changing the way that foreign dependencies are integrated into the rust toolchain (i.e. moving from a rust-centric system like Cargo to something more general like buck2)
Worth noting that the JVM ecosystem doesn't have this issue because there is no notion of installing dependencies, and the package managers are just downloaders with nothing else. No build.rs equivalent.
To solve the problem of native/C dependencies, they just bundle pre-compiled libraries as data files.
It may or may not be a problem depending on whether it's actually used, which is still better than "gets you instantly the moment you compile the app". A lot of codepaths and even whole libraries aren't invoked just because a program depends on them.
It doesn't really matter anyway, because nobody is reading anything in their dependencies before it gets downloaded. Malware can also be hiding in plain sight in source code, as this attack and many others shows.
ABI stability is significantly different in a VM language to a native language, and reified generics necessarily require source compilation. Binary-only development for Rust would be exactly as 'sort of not really' as C++, for exactly the same reasons. (There is also no notion of 'installing' dependencies in Rust, and build-time code being malicious isn't much worse than runtime code being malicious.)
The vast and overwhelming majority of build scripts are building C code, so the other solution is to move away from integrating with C dependencies to native Rust dependencies, in which case adding friction to build scripts would be less noticeable.
Just denying write access outside the build directory and denying network access would go a long way and won't break pretty much any well-behaved build systems.
> Just denying write access outside the build directory and denying network access
As the link you posted mentions, you need a tiny bit more than that: you also need write access to the temporary directory (/tmp and similar). Many build tools temporarily store files there; for instance, unless things have changed since I last looked, if you don't use the -pipe argument the C compiler stores its temporary intermediate files (preprocessor output, assembler input) there.
One of rust's strengths is it's ability to interface relatively easily with existing c code without having to rewrite absolutely everything in rust. I don't think that is something we want to give up.
build.rs changes nothing about how easy it is to integrate with C. What does simplify: figuring out how to supply library you need at build time.
Which is the result of how bad dependency managment is outside (i.e. DLL-hell).
Pretty much all other use cases of build.rs can be sandboxed. Well, there is sqlx that wants to connect to database at expansion time to compile check-queries (yew).
sqlx at least has the (optional) offline mode, where you "cargo sqlx prepare" once (which wants access to a db) and then you can build in offline mode which typechecks your queries against local files.
Although I suppose that's still doing a lot of shenanigans at compile time. It could be sandboxed pretty well (theoretically). I'd hate to give it up completely though, getting a compile time error when SELECT query params or return values have type mismatches is extremely nice.
sqlx offline mode being opt-in instead of default is what bothers me. You know what else can validate that your queries return what you expect? Integration tests. Shoutout to sqlx for #[sqlx::test] though.
I also wish that it would be the default. Integration tests are fine, but they aren't compile time. Elevating them to compile time and using an LSP enabled editor makes it just underline SQL errors before I've even had a chance to run a manual compile let alone a test...
LSP diagnostics is actually what made me switch from compile time check queries. Whole "is db up? are migrations applied?" dance tired me pretty quickly. Its fine if you use sqlite, but anything else gets annoying.
Sandboxing for build scripts can't work properly. If you sandbox too much, some necessary stuff can't be done. If you sandbox too little, it has no practical value.
As an easy start, how about letting build scripts read /usr, read and write a temporary build directory, have some /tmp scratch space, and be allowed to write its final output artifact. No network and otherwise isolated from the rest of the system.
I would argue that, if a build script doesn’t work in the setting, then it doesn’t deserve to be installable by a default cargo command.
Cargo is a cross-platform tool, so when it ships a sandboxing solution it will need to be a cross-platform solution, and because this is a security feature it needs to be bulletproof, so no half-measures like Docker. Something like a WASM runtime might fit the bill, though that will be much easier to get working for typical proc macros than for typical build scripts. If you only care about Unix, then you can do this yourself today by building code in your sandbox of choice.
> Cargo is a cross-platform tool, so when it ships a sandboxing solution it will need to be a cross-platform solution
This seems like an excuse, not an actual objection.
Linux can do seccomp or Landlock or gVisor or a combination. Seccomp and gVisor need no privileges. Windows has its internal weird mechanisms. Mac has sandbox-exec.
Cargo could easily pick an appropriate sandbox for each major platform and ship it by default.
> If you only care about Unix, then you can do this yourself today by building code in your sandbox of choice.
This is ridiculous. The sandbox should not have network access, but cargo needs network access to download the package in the first place.
So let each build script define its own level of sandboxing and then users can determine whether they are okay with that level or not, e.g. `cargo build --sandbox-level=...`
Rust, like C, C++, and every other systems programming language, is all about giving users the power to make mistakes. The philosophical difference when it comes to Rust is simply that it tries to force the user to flip off the safety on the gun before letting you shoot yourself in the foot. A Cargo config option letting people opt-out of sandboxing would be fully in line with Rust's philosophy.
Sandboxing just build.rs would only be
be a minor inconvenience for the attacker, nothing more. The attacker can always as easily compromise the binary you build and as soon as you run it (e.g. in a test) you are owned.
It would be a big pain for many that are in the unfortunate position to really need build scripts, though.
I imagine it would be sandboxed by default with an escape hatch to run build scripts outside of the sandbox with user verification. It makes people stop and think about what’s happening. Not perfect, but it does help. When working on JS ecosystem projects I manually approve build scripts and spend some time researching dependencies with build scripts to see if I can avoid running the build script. Some people will ignore it and run everything, but it’s a huge step in the right direction to make it operator-decided.
It would be more than a minor inconvenience. I can handle sandboxing my tests and production infra, but I can't handle sandboxing build scripts because I don't own that code in any sense.
At the very least, it wouldn't be overly onerous when adding a dependency that requires a build script to require an opt-in via Cargo.toml, e.g. `build-script = true`. You'd make it viral so that any transitive dependency that requires a build script would affect its parent, then add the key as defaulting to `true` so as to not break backwards-compatibility, then switch the default to be more restrictive over a new edition. (This same key could be used to prevent proc-macros from having arbitrary system access as well, where by default proc macros could be compiled to WASM and run in a WASM sandbox and treated as pure functions.)
I mean sure, but anything the build script could do, the build artifact could also do, That is to say, if you don't trust your source why do you trust the thing it compiles into?
cargo add + rust-analyzer instantly executes build.rs before you have a chance to audit the code.
Cargo, please PLEASE give me a way to disable third-party build.rs and whitelist the ones I need. And please loudly mark any update that adds a build.rs where there was none before.
cargo-deny can audit build scripts, but unfortunately not prevent execution of malicious build scripts exactly for the reason you gave. It could still help if you only ever use cargo add and update in a sandbox.
Safe-ish recommendation: Only add well-known, trusted crates. Failing that, treat unknown crates as malware or containing malware dependencies until proven otherwise. Test untested tools/crates in a VM/dev container to be sure they work properly before trust them. Use Mark I eyeball and Mark III brain too. :)
Audit and limit crates (cargo-deny &| cargo-crev, && --offline) until tools exist to audit build.rs safety semi-automatically ($$ safeguard.sh maybe).
The root problem is two parts:
1. crates.io doesn't do mandatory curation. Lack of curation is fail. It's time-consuming and costly for reviewers without a doubt, but so is letting an ecosystem gain maximum entropy (go to shit) by Tragedy of the Commons depending entirely on the honor system. Name squatting, low-quality, unmaintained, typosquatting, and malware are the consequences of too much self-service / semi-self-service freedom.
2. Many, many cargo subcommands are over-eager to run build.rs because it assumes trusted crates:
In an untrusted/uncurated crates world and a build.rs exists or exists in a selected dependency, it shouldn't run at cargo-add time (unless it must). If it exists, on first run, it should be presented to the user in a viewer for manual review unless a magic CLI flag/env var is specified to accept it.
These 2 factors combined appear to create a Swiss cheese holes failure mode for running arbitrary crate `cargo add`. I have confidence a suitable add-in workflow &| standard command &| repository workflow will be adjusted to reduce the attack surface of the ecosystem.
I think it would be a good start if crates at least had to opt in to a build script, and adding one later would require permission from crates that depends on it.
The vast majority of crates don't need build scripts, so it is vaguely feasible to audit the list of crates you use that might need them.
There's no good reason a proc macro can't run in a no-IO sandbox by default. None. Doesn't require a language change. Doesn't require some microvmcapabilityeffect BS. It requires looking people straight in the eye and saying "no" when they complain about needing to prompt for privileges.
Rust suffers from the same faults as the JS ecosystem. Any significant crate imports hundreds if not thousands of dependencies. The probability that one of the authors gets targeted by AI-assisted attacks is just too high.
Also most of these dependencies provide a breadth of features that the end package does probably not need.
My experience has been that it has a major advantage, in that freeze + offline actually work properly. You can collect the dependencies you need once, put them in version control and never ever talk to remote registry again
A language without a large stdlib pushes this that functionality into (transitive) dependencies. I hope more language will adopt batteries included approach.
No stdlib will ever include all the "batteries" you need. And large stdlibs have their own set of problems, like stagnation, tying the library version to the language version, backwards compatibility garantees preventing evolution, and either needing a wide range of domain experts to maintain it, or having developers maintain components they don't have a deep understanding of.
I think the sweet spot is having an ecosystem of "blessed" libraries that are reviewed for security and quality, but are versioned independently of the language, and maintained by subject matter experts.
Doing software development outside of strict containerization, at the very least, looks increasingly prone to disaster.
Yes, we can argue about the culture of package management (as some of us have with especially npm from day one), but it's done, and your colleagues or AI sidekicks cannot be trusted not to download whatever and try to build and run it. All you can do is limit the effective blast radius.
1. There isn’t a single universal standard for sandboxing across all the different platforms that are supported by Rust.
2. Even if there were, if you’re compiling untrusted code then why would you trust the built output?
If you’re building the create then I’d argue that any preventative steps afterwards is akin to closing the barn door after the horse has already bolted.
> rust-rfcbot commented 2 days ago: This is now entering its final comment period
I wonder if this is not in part the same situation as we arguably had with the xz compromise: some imminent change would have made the attack harder (in xz's case, IIRC a change to systemd to dlopen the compression library instead of directly depending on it), and the attacker rushed before the opportunity window closed.
It provides sandboxing for arbitrary CLI commands using Seatbelt / SBPL on macOS and Landlock LSM + seccomp-bpf on Linux. You can use it to protect local dependency builds, or integrate it into GitHub Actions to add an extra layer of protection to CI.
Feel free to give it a try — I’d love to hear your feedback.
We need effect based languages now. It's the only way to guarantee policies like no network, no file access, no unsafe code or FFI for a library before it even compiles. If anyone from epic is reading this please give us a timeline for open sorucing the Verse compiler.
In the mean time I think it's possible to hack Cargo and run all build scripts in a microVM. The blast radius will be limited to malicious code in the binary instead of uploading all your CI secrets and deleting the whole hard drive.
I think this is an attractive chimera, but will never be popular and will only be used much in specific domains.
The thing is, trust is a massive optimization. If you believe code from X is competently made and not malicious, you can accept that code without any additional overhead. (Building trust is a significant effort, but trust begets trust, so there's a resonable path from starting small to scaling.)
It's not that these mechanisms that require less trust (none of them eliminate it, BTW, when you examine them) can't work. It's that they'll inevitably be outcompeted by trust-based systems (in most domains).
E.g., compare the effort of creating a new, full featured, general purpose language, promoting it, building and maintaining all the tooling and documentation around it to support it, etc to the effort of creating and maintaining a curated package repository of trusted packages (and/or adding trusted code to extend the standard library) for existing language X.
(It's not clear to me you could provide enough value from a new language focused on effects that you could convince any but a few from adopting it at all, even if it dropped from the sky fully formed with all tools, docs, etc. in place from day one.)
> It's the only way to guarantee policies like no network, no file access, no unsafe code or FFI for a library before it even compiles.
It's not the only way. You can sandbox processes. I think sandboxing is the much more reasonable approach, because in the end of the day, there are still closed-source software products where you can't demand that the manufacturers use certain language safety features.
Sandboxing the process only works well when the malware requires more capabilities than the software itself.
So if your software needs to make HTTP requests and read the filesystem then the malware will be able to make HTTP requests and read the filesystem, which is enough for a ton of malware. Sure, maybe you can limit the directories it can access a bit and possibly some sort of network filtering, but it isn't a silver bullet.
Capability security in-language would make a huge difference, because the more granular you "sandbox" the less likely it is that the compromised component has the access it wants. For example if the HTTP client library is compromised maybe it can't access the filesystem so can't steal your cookies. Or maybe like in this case no permissions were needed at all and despite the process having enough capabilities for malware this malware can at worst return bad values and try to chain this to an exploit which is far more difficult than just running it itself.
(as I and others have mentioned in the thread): The attacker can just move the malicious code from build.rs to lib.rs (ie. build-time -> test/execution-time).
Then the problem is the language, as the grandparent observes.
I don’t even understand what a capability based language is, but presumably closed source software wouldn’t apply here? Is it common for closed source software to give the source code to customers to build?
> Is it common for closed source software to give the source code to customers to build?
Yes, somewhat common. I work in embedded, and where a supplier gives us the ability to build their code it is a lot easier for us to deal with the next time we need a new build from them.
Closed source software could apply too, if the execution model supports it (e.g. WebAssembly). Just because the format is binary and "unreadable", its permissions (accessible functions) don't have to be
I think the only language that's used in production where malicious dependencies can't do arbitrary effects is https://roc-lang.org - but it's a pre-0.1.0 language and as such its production usage is extremely minimal for now.
This is why I'm building this language. It's capabilities based. Build scripts can't just do whatever the hell they want to, everything is auditable, and it layers its sandboxing.
It's for fun and anyone looking should understand that this is AI driven building with human driven design, but the goal is to demonstrate.
Can you explain how an effect based language helps here? Also to note filesystem access can allow network access on many systems and also the other way around.
All languages in mainstream use are imperative and at least somewhat object oriented, other than SQL. But many of those don't have what it takes to implement capabilities (too flexible/dynamic).
First we need capability-based OSes, like we should have had decades ago if worse-is-better hadn't stuck us with Unix. I don't need to care whether or not a program was written in a capability-aware language if the OS fundamentally takes care of that for me.
HarmonyOS is a capability based OS. I don't know why anyone in China is still using Rust. The biggest social engineering hack was for the Rust team to convince developers that it is a safe language.
1. seL4 exists. I haven't seen much more formal and rigorous than it. Efforts have ported the Linux kernel and RTOS-things to it.
2. Add syscall(s) to POSIX, which encourages *nix adoption, with an API like the following:
- Drop capabilities for thread/process except allow list.
- Test if current thread/process has all listed capabilities.
- List should be able to represent capabilities and their attributes, much like a single unveil() OpenBSD call, maybe as an array of/pointer to struct.
- (Program crashes if syscall capability is missing.)
Why does every build have the ability to run arbitrary code by default in the first place? That seems like something you should opt into only under very specific scenarios and even then it should probably be built around a WASM sandbox or microVM (as the parent suggested).
Because many many builds end up doing something just a little weird that the build system cannot handle by default. Ideally the build system would be fully property/dependency based and so it wouldn't have to run arbitrary code. In the real world everybody has something weird about their build that the build system cannot make work.
This is partially because for nearly every project the build system is something they need and don't care about. When they need something weird they hack just enough to make it work and never ask "how should the build system change so that this was a property instead of running code", and thus build systems are slow to improve. In a number of cases the build system did have a way to do that thing, but the person didn't know about it.
There are also a lot of code generators out there. I have yet to see a large project which didn't have their own code generator for something specific to their project (protobuf is an example from Google that has escaped and become useful elsewhere, but there are many others that are specific to one project. Yacc is from the 1970s, and stands for "yet another compiler compiler" - implying the idea was already common 50 years ago). You cannot have/use these useful tools without running arbitrary code.
I don't mind code execution as even make has shell scripting embedded, same thing with meson. My issue is with network access to download random stuff without it being declared somewhere and signed.
I'm using OpenBSD and the ports system set up two users `_pbuild` and `_pfetch` for building packages. The first one is for building and the sample `pf.conf` (firewall) forbid it from accessing the network. The second does fetch the files , but it's declarative with every files listed and signed. Even for languages like go and rust.
Fetching files while building is the bad idea there.
Good point. Many build systems also want to be a package manager, but the two are and should be separate. If you have a simple problem it makes things easy to combine them into one. However they need to be separate anyway, both for security and also to make other weird situations easier.
MicroVM this, effects that --- can we discuss security without needing to invoke bay area buzzwords?
The idea is "least privilege", and we desperately need it in computing. The precise technical mechanism we use to achieve it is less important than committing to the idea that a dependency doesn't run with full privilege of its host program and an install script doesn't run with full privilege of the programmer.
The original Linux seccomp is old enough to drink. It's always been possible to do things like expand macros in a no-IO environment. Nobody's bothered to do it over the past two decades. Why would anyone bother in the next two decades?
The technical mechanism is exactly the issue to figure out, there's a reason why projects don't have this and it's because different implementations have different tradeoffs.
No, the reason people haven't been doing this -- and we've had technologies for ages -- is that it's a huge pain in the ass, a "tax", that it's hard to get developers inside a company to pay, much less participants in open source ecosystems.
Look at how snap, flatpak, etc. provoke people to just turn off security rather than deal with breakages.
A new language won't help because the problem is social, not technical.
FFS, you can't even get people to use filesystems via intermediate objects in today's languages. People think your language is broken if you don't have an ambient open(). You don't need a new language to enforce capability discipline. You need to whack people repeatedly with a cluebat until their laziness and brain damage abate.
You can't simultaneously say that we've had the tech for years and also say that it doesn't work. We have not had the tech for years. x-plat sandboxing is extremely difficult, especially in a way that's performant.
Sandboxing almost always involves having a dedicated, privileged service, due to how operating systems have designed things. It requires platform specific code.
This is a technical problem and a social problem, there's zero reason to believe it's just one.
The technology works fine if you use it in a disciplined way. The problem is that people don't! You're necessarily going to break programs if you put them in restricted environments. That's not the same as the restriction technology not working properly.
> Sandboxing almost always involves having a dedicated, privileged service, due to how operating systems have designed things.
Or just invoking bwrap. Or sandbox-execute. There's no performance overhead. Complexity is minimal. Just read current Codex source code. It's not so bad. People have been making these sandbox tools for AI agents for years. Just need to apply sandboxing to all domains.
You want some kind of silver bullet that makes code "safe" without anyone having to change anything? To prompt for nothing? To use no IPC portal? That's not happening.
The people saying we need a new language or something are half right. We don't need a new language. A pure library solution is fine! But people do need to do work to make an ecosystem based on capabilities and least privilege to work. And we need to tell people who refuse to do this work to go to hell.
A new language is neither necessary nor sufficient. What we need is new set of balls.
> The technology works fine if you use it in a disciplined way.
Then obviously we have to discuss the implementation!
> Or just invoking bwrap. Or sandbox-execute. There's no performance overhead. Complexity is minimal. Just read current Codex source code. It's not so bad. People have been making these sandbox tools for AI agents for years. Just need to apply sandboxing to all domains.
Tools like bwrap literally could not have existed until very recently without also requiring suid/ privileges, and even today unprivileged user namespaces are not universally enabled. This is why the implementation matters. No, unprivileged sandboxing has not been around for years, especially not x-plat. bwrap is a perfect example of what I'm talking about, great reference - it either requires root or it requires unprivileged user namespaces and it's not x-plat. Great.
> You want some kind of silver bullet that makes code "safe" without anyone having to change anything? To prompt for nothing? To use no IPC portal? That's not happening.
I have no idea why you think I've said this, I'm pointing out that the implementation and technology matters deeply.
> The people saying we need a new language or something are half right. We don't need a new language. A pure library solution is fine! But people do need to do work to make an ecosystem based on capabilities and least privilege to work. And we need to tell people who refuse to do this work to go to hell.
Again, you can't say "this technology has existed for years! You can do it in a library!" and then say "but they don't!". You misunderstand the complexity and assert that it's merely a matter of will. It is both.
Browsers have been sandboxing their tabs since 2008. Don't sit there and tell me that bwrap, sandboxing, whatever is some kind of recent innovation with uneven distro support.
And yes, bwrap needs root or namespaces. So what? So does sudo. Setuid binaries are as old as time.
You know what's also been around and doesn't require root? Landlock. Have you heard of it? Has anyone? There's a huge PR and visibility gap in this space. We don't have to invent some new thing. We have to get people to learn about what we already have and use it.
"Oh, the technology works, but it need setup! Oh, my distro doesn't enable it! Oh, it doesn't work out of the box with my code!"
This is a social problem. It's not a technology problem. It's people not wanting to do the work, not understanding what already exists, dismissing solutions based on non-problems (like bwrap needing privileges on some systems), and in general adopting an attitude of "no", not a can-do problem solving stance.
And we're supposed to solve this problem with let another precious little effects language or another fucking MicroVM environment? FFS.
We need unprivileged sandboxing!"
"Use bwrap"
"Nooo, that needs setuid root or file caps!!!1!1"
"So the problem is that bwrap is part of the TCB?"
"Yes. We need sandboxing that relies only on unprivileged code like the Linux kernel"
> Browsers have been sandboxing their tabs since 2008.
Yes, it's an amazing feat that has cost billions and led to major features like seccomp v2, ptrace sandboxing, etc. Do you know the history of browser sandboxing? It's pretty complex, a major technical feat.
> Don't sit there and tell me that bwrap, sandboxing, whatever is some kind of recent innovation with uneven distro support.
I can tell you that unprivileged sandboxing is not only new, it's ongoing and nascent work and not commonplace at all.
> And yes, bwrap needs root or namespaces. So what? So does sudo. Setuid binaries are as old as time.
These are massive footguns and issues for deploying code... You don't get that? Again, implementations have tradeoffs.
> You know what's also been around and doesn't require root? Landlock.
lol it is SO NEW what the fuck dude?
> Have you heard of it? Has anyone?
Yes, I use it!
> "Oh, the technology works, but it need setup! Oh, my distro doesn't enable it! Oh, it doesn't work out of the box with my code!"
> This is a social problem.
Yeah these sound like social problems if you have literally no idea what you're talking about lmfao. IT's BOTH.
Your ignorance is blatant and your position is dead in the water.
From the article: “[for Windows victims,] the [malicious] build script [fetches the attacker’s remote payload,] writes [it] to %TEMP%\rust-setup.ps1 and starts [it] through a VBScript launcher under wscript.exe, with a comment in the source explaining why:”
And the comment is:
// ShellExecute via WScript escapes Cargo's job object; spawned children otherwise
// keep the build script (and `cargo build`) waiting until they exit.
So the malicious build script has a helpful comment (???), written in a familiar “terse nouns verbing” style (!!!).
Would it be gauche to speculate? Maybe some script kiddy sweet-talked Fable into dropping its safeguards, or maybe Anthropic is doing a training run for Fable 5.1 and the air-gaps aren’t gapping.
I mean, yes, update your dependencies. But probably after a week or so after they've been release and tested by the first penguins willing to jump into the ocean.
Holding periods for new updates are a good idea, but I would also prefer having a tool that could provide diffs of the entire project state pre- and post-update; including transitive dependencies being added or removed. Git diffs won't show you that information, by design.
There's a reasonable solution to this. Tools like `cargo-audit` can tell you if a crate has security issues and guide you towards an update. You don't have to blindly update all dependencies this way.
Nothing to do with Rust as a programming language, but meanwhile in the world of [language without a package manager], a monkey puppet glances awkwardly to the left.
The rust ecosystem is going to be hit by malware just like the NPM ecosystem. I have been saying it for years. They made the same mistakes or even worse mistakes because all it takes is a compromised serde to take the entire ecosystem down.
Here's a suggestion, do you think it could be this simple? If you write a package-manager, don't execute any of the downloaded code in an automated fashion. No hooks, no build-time metaprograms. If your language _needs_ metaprogramming to function, it's a huge secondary issue that I don't know how to solve. You can try to make a meta-program annotation which disables side-effects, but the metaprogram ultimately must write memory which expands the amount of code generated. All risk is introduced when the person downloading cannot preview the content in a safe place.
Unfortunately it doesn't solve the problem because malicious code can still end up compiled into the user's program which they will promptly execute, possibly in production...
Supply chain attacks are not a problem that can be solved by a single silver bullet, however the biggest benefit comes from a combination of minimum release age + fresh 2FA required for every publish + automated scanning. This makes it considerably more difficult to pull off a supply chain attack and should be the baseline security for all package managers.
The only safe solution is to review all the code's changes (or trust someone else to do it).
Forbidding compile/build-time shenaningans is trivially bypassable and has already been bypassed in the NPM ecosystem by just making the library code itself (not the build scripts) malicious - eg. do the bad thing when the code is loaded/tested, assuming the language has static constructors.
Summary: it's cultural. Rust likely inherited the practice from Nodejs, who inherited it from Ruby. I think in Rust online spaces in particular there is also this undercurrent of "you're not smart enough to use certain parts of the language, so download libraries that handle that stuff for you."
The language also takes backwards compatibility very seriously, anything that goes into std "must be maintained forever". It is also argued that a large stdlib means the maintainers have less time to work on the language itself.
I understand not every language can have Go's amazing stdlib, but I would much prefer Pyhton's approach where every now and then some package/function from the stdlib gets deprecated/removed. Rust's 3rd party ecosystem is the worst thing from the language, worse than the compile times.
i'd like to welcome you to the hell that is c/c++ dependency management. Make? cmake? qmake? conf? autoconf? configure? autotools? submodules??? AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
I'd just select a tool, vendor the libraries I need to the codebase and call it a day. Did these for Eigen, liboption++, Catch2, Easylogging++ (now archived and not maintained anymore, sadly).
Build a simple makefile, and you're off to the races.
None of those (other than submodules which is just vendoring) does dependency management. They are more configuration management than anything. C, Python, Perl (and maybe ruby?) relies on flag switches and environment variables to find all the necessary files and modules for compiling/running a script. Cmake and autoconf just configure those.
With NPM and Rust's focus on project's level dependencies, there's no longer emphasis on API stability. Instead we have breakage every months, forcing everyone on the upgrade treadmill. It's easier to audit C library because they focus mostly on security updates instead of redesigning the API for the nth time.
The "npm-ness" of Cargo (centralized and standardized dependency management used at every opportunity) is generally pitched as one of the primary developer experience advantages over C++.
There are two reasons you might want to use a crate, The first being that you want to use a good solid implementation that you know someone has spent more time doing and works better than almost any solution you could integrate. The second is you don't want to spend time implementing that.
Writing macros in rust is a pretty horrible experience but it's not difficult
Rust is actively incorporating more functionality into the stdlib. The functionality of this crate has been in std since 2024. The ecosystem is just slow to update (not everything is maintained, etc).
The languages that have a poor standard library support have this issue and other languages encourage you to import tons of libraries to fix the problem.
This is why Javascript and Typescript suffer from this the most and has little to nothing to do with "popularity" and likely 9/10 of these npm packages import an external library.
Golang on the other-hand is just as popular and has a stronger standard library which people build against and it is encouraged to use its standard library rather than rolling your own or importing another package to solve the problem.
Rich official standard library vs "import tons of libraries" are not the only two options.
Java's standard library had arguably also been poor for a very long time and "import tons of libraries" just had not been practical for most of that time because the tooling and ecosystem for that did not exist yet.
The solution was apache-commons and guava. Two large libraries with everything the developers heart desired and well maintained by large organizations.
For Rust be probably will never have anything exactly like that because requirements from no-std development to fully fledged backend service are too diverse, but there is still room for a small number of well
maintained backed by reputable developers convenience libraries in my opinion.
Namespacing. Mandatory namespacing. It's an essential prerequisite for any kind of reputation building and managememt, before you even get into the gritty technical or security details.
And it has to be mandatory. Top-level package names will always have more cachet. Developers are suckers for good package names, literal or imaginative. Plus it helps address, but by no means completely solves, name and typo squatting.
I understand people and groups can run their own crates.io-like repository, but that's a tangential aspect. Even if this were ubiquitous, you'd still want mandatory namespacing. You want provenance, or at least intended/nominal provenance, to be as transparent as possible, not implicit or buried. By no means a complete solution, but an important foundation for better technical and culture patterns.
I don't want to see a large standard library for Rust. If something is added to the standard library, then it is very difficult to change it afterwards because backwards compatibility.
It would be better to have blessed crates in crates.io. The Rust core team would release or audit them. If the blessed crates need breaking changes, it can be done by increasing their major semantic version number. That can't be done to the standard library.
Actually, there could be a "trust" level for crates: 1. blessed crates by the Rust core team, 2. trusted developers, 3. untrusted developers. Or something like that..
Go has a stronger standard library for certain use cases like basic CRUD web applications, but a lot of the Go standard library is also extremely low quality (flag, container, image, json, log, math, path, regexp, sync, time). Many of these aren't usable outside of toy use cases and have weird edge cases all over the place. Over time many will probably get new incompatible versions just like json. The container package in Go is the worst collection library in any mainstream programming language by a huge margin. It's actually astounding how bad it is.
> a lot of the Go standard library is also extremely low quality
Now adjust your definition of "a lot" to include all other languages instead of apparently arbitrarily deciding that "a lot" means at best 10 and pretending that "weird edge cases all over the place" isn't normal.
I also wonder what your issue is with most of those, especially since e.g. regexp is excellent in the context of this thread - it runs in linear time, so it's not possible to craft a malicious input that will make it slow down to a crawl.
> Over time many will probably get new incompatible versions just like json.
What a bizarre statement to make.
FYI: "The encoding/json package is now backed by the v2 implementation.". You get all the benefits possible from the v2 implementation for free while maintating backwards compat and if you want to upgrade to something better, v2 is right there.
Do you have some magical suggestion how this could have possibly been handled better? And no, not having json in the stdlib isn't a viable path, that's just a cop out. But I guess since e.g. Rust and Java don't have json in their stdlibs at all, you can't say their json packages are bad, how clever and smart!
Another frequently made suggestion straight from la-la land is just writing perfect code from the get-go. Brilliant, can't believe nobody's ever thought of that.
Meanwhile in the real world, you've been able to reap the benefits of the solid encoding/json for over a decade now, and with v2 you get some nice free backwards compatible improvements and have a clear path to upgrading to v2. Perfectly handled IMO.
I don't think this is true. People like to scapegoat the JS standard library but in reality most JS implementations have pretty featureful standard libraries, especially browsers. I've never felt the need to use any third-party libraries in the javascript projects I've worked on (except for maybe `ws`).
I don't think it's just that, though I agree they are clearly correlated.
The reason I don't think it's a sufficient explanation is that there is a clear history of large, 3rd party libraries being created exactly to supplement poor standard libraries. C++ has Boost, Java has Apache Commons (though Java also has a pretty huge standard library), arguably we could even say C has Posix/Win32/Cocoa.
I believe there is some deeper cultural reason why certain language ecosystems coalesce large utility libraries, while others prefer myriad tiny dependencies.
The important difference IMO is in the small vs big libraries culture.
Lots of languages have a bad stdlib but don’t fall into the trap of having thousands of micro libraries.
The reason people do it is because it brings clout and money. Just look for articles defending micro libs: the popular ones are by people who make a living on donations, due to maintaining 1000+ packages.
And collaborating in larger libs/stdlib is hard. Plus: Rust, Node, all have a lot of visibility.
You need a good stdlib culture to avoid it (like Go did).
The reason is simple: the package manager works well. Helps if the package manager is standardized and there is a "default" registry for open source projects.
I remember the days where I had to manually put the Spring .jar files into my project. No way I am doing that for 100s of dependencies.
I've never seen a non trivial Python or Go project without external dependencies. The dependency tree of comparable Go and Rust projects seem comparable, IMO.
Pytorch has 9 transitive dependencies for cpu-only execution, or 29 to bring in cuda.
candle, the most popular rust ml library I found in a cursory search, has 119 for cpu only, and 150 to bring in CUDA.
I guess it's taste whether that's comparable
this would have been a way more satisfying dunk if nvidia hadn't split the cuda functionality needed by pytorch into 19 (!) packages on pypi but such is life.
I expected somebody to pull out an example "disproving" my point. It wouldn't be hard. I can point out rust projects that have a ton of external dependencies, and then point at similar projects with very few.
But a Python example doesn't really count, in my mind -- Python is pre-GitHub so tends to have small numbers of large external dependencies, like C++ and other older languages.
Unfortunately Cargo doesn’t have security controls in place to prevent these kinds of attacks. For example pnpm has controls to allowlist install scripts for dependencies and will warn about new install scripts (without executing them).
Compromising the code that is then most likely run in a test instead of compromising a build script is just a very slight inconvenience for the attacker.
I share the dislike for arbitrary build scripts but restricting them will not help the supply chain issue in a significant way.
Also there are several ways to control build.rs execution in the Cargo ecosystem as well, for example with cargo-deny.
You’re right about attackers being able to change runtime code.
pnpm does have some other features to prevent supply chain attacks, so there is still something to learn from other ecosystems. For example pnpm has a cooldown period for new dependencies and can prevent trust policy downgrades (eg new version published without build provenance where older versions did have it). See https://pnpm.io/supply-chain-security
Disallow lists are ineffective - better to disallow by default and require opt in.
Also, crates.io can defer serving up newly uploaded scripts that have a new build.rs / proc-macro dependency and warn publicly that a version introduces it.
Restricting build scripts 100% will help mitigate the impact, just not if you only deny it once. And they can develop other things like sandboxing for build scripts by default and escaping that to be the exception that has to be explicitly allowed.
Are there any plans to more seriously develop the standard library in Rust? Or is the plan to remain in this status quo where users of Rust import nonsense and the dependency tree explodes (or users are forced to invent their own wheel?).
Are there any comparisons between the state of the stdlib in C++ vs. Rust? I’d think that would serve as an excellent jumping off point to start chipping away.
The standard library is not versioned, so any API*/behaviour there must be maintained forever. When you put it in a separate crate, you can make better interfaces or be displaced by a better crate, while old code still compiles against the version it was written for. So even code effectively maintained within rust teams (hashbrown, rand, regex) may be in separate crates, std is specifically for OS abstraction and a shared type vocabulary.
That doesn't mean you jump to importing nonsense or trivial dependencies. The top hundred are efficient, well-designed crates of the quality you'd expect in a large std like Go or Swift, sometimes even higher as people can write better implementations that they otherwise wouldn't (or wouldn't be used over std). And those languages make breaking changes! C++ is mostly stable, so it's full of junk like <regex> or just unordered_map. Rust managed to wholesale reimplement HashMap 6 months after SwissTable released, like it is also easier to express this level of encapsulation, but that's part of countless design/interface decisions made deliberately to not constrain forwards compatibility, including a smaller std.
However, that's no excuse to have a worse developer experience in this area. We need better tools to vet and communicate the quality of a crate and its supply-chain, like community-curated or even additional org-maintained crates, and maybe a handful delivered precompiled in the default rustup distribution which can change over time. I don't think they need to be added to std itself though
The idea that Rust has a small standard library is a weird meme. Rust has an enormous standard library. Look at any Rust release and you'll see dozens of new library APIs added, and consider that Rust has about nine releases a year, meaning that Rust adds over a hundred APIs per year, and has consistently for the past ten years. Rust frequently adds things that obviate popular crates, most recently the cfg_if crate was made redundant by the new cfg_select macro. And half of every dependency graph that people wrong their hands over are actually first-party external crates provided by either the Rust organization itself or by known contributors to the project. When people say that Rust has a small standard library, mostly they seem to just mean that it doesn't include a webserver.
JSON: there are a ton of different ways to do serialization and deserialization, each with their own tradeoffs, and serde (the most popular) is far from universally agreed upon. The same goes for JSON specifically, there are many different serialization formats with different tradeoffs.
regex: Owned by the rust-lang organization already. You get the benefits of trust (if you trust std, you trust the rust-lang organization anyway), but without the issues of being in std (backwards compatibility and bloat). The only problem I see with that is that BurntSushi is still the owner of the package and as such can still publish new versions on his own (AFAIK crates.io currently requires at least one user owner, but that's something that can be solved by improving crates.io permissions).
RNG: rand is still evolving, with breaking changes half a year ago. Preferred generators tend to change over time, so I don't see those getting into std (remember, it has to be maintained forever!). I could see the interface (traits) getting into std, but I see no advantage to it: it's maintained by rust-random, but even if you wanted it to be maintained by the same authors as the Rust language (let's say you trust them more than rust-random), you could just move it back to rust-lang-nursery or rust-lang.
CLI arg parsing: not simple at all! The community's preferred solution has 70kLOC (with support for so many features), but there's also argh, pico-args, and others, each with their own tradeoffs.
> That's why projects end up with 100s of crates, sometimes 1000s.
Also because libraries are split up in many crates. For example, regex is split in 3 crates: regex, regex-syntax and regex-automata, and depends on other crates from the same author, such as aho-corasick.
All that said, there are many crates, such as algorithms like aho-corasick, that I think could me moved into the rust-lang org, like regex was.
Cargo (and PyPI) is undeniably better than NPM, which is just shockingly bad for cultural reasons. Yet it's not safe, and it's subject to the same class of exploit, as we're seeing.
Indeed, the solution is to get away from the wild soup of author-managed dependencies and go with something with an audited collection of software that is maintained by separate human beings from the known-vulnerable hackers writing the software.
And indeed Go and .NET and Java all qualify. But the gold standard here is Debian and all its downstreams.
> the solution is to get away from the wild soup of author-managed dependencies and go with something with an audited collection of software that is maintained by separate human beings from the known-vulnerable hackers writing the software.
I think we might be able to crowdsource audits. At least in the Rust ecosystem I'm confident that this is feasible with the right tooling.
Oh, so it's not only "JavaScript bad and npm bad". Apparently, if your language uses third party dependency registry, you are prone to malicious code, regardless if it's Javascript or not.
Use containers for development. And reduce the amount of third party deps you import into your projects. This is only going to get worse.
Pulling in lots of dependencies creates this kind of risk regardless of the ecosystem. That being said in the JS/NPM world you tend to have a LOT more dependencies (especially indirect ones) than other languages. I saw someone do a cursory analysis and JS/Node projects tend to have 5x the number rust or ruby projects.
This is really a cultural problem not a technical one.
If it weren't a registry it would be ./configure scripts and makefiles. The issue is that sandboxing technology is kinda shit (especially x-plat) and languages don't build it in by default.
> The genuine arrayref and append-only-vec crates are maintained by droundy, whose account appears to have been compromised.
That name looked familiar to me; I believe it's the same David Roundy who was an academic at Reed College who wrote DARCS, which is version control software. I used DARCS when I started grad school around 2010 before switching to git.
I would like to know too, but the author's Github account seems to have been deleted, and the crates.io releases have been completely deleted too (not just yanked) so it seems impossible to view the file :/
The author proposes to reinvent the wheel and depend on stale dependencies, which are EOL. Neither of that is an acceptable solution with LLM-based Agents being able to produce exploit(ExploitGym) chains in minutes from known bugs. The other issue is that Rust's forces the user to provide more information and APIs are usually kept generic for systems programmer, so standardizing things is not as straight-forward compared to Go where you can assume a memory-management, a virtual thread runtime and mostly ignore dynamic dispatching.
I don't think anyone is saying that you should depend on stale dependencies. You should not hand off your responsiblity to untrustworthy package managements services though. Evidence of the failure of automated package management is the near constant shai-hulud attacks as well as the originating post here.
The fact is that your dependencies are your responsibility. You should be staying up to date on what you depend on and their vulnerabilities. Most vulnerabilities affect specific code paths which may not be present in your code base and do not require mitigation. These are decisions you can make by understanding what your dependencies are instead of offloading to automated package management systems.
NPM has been the posterchild of these types of issues but the dependency stack I saw everytime I compiled even simple programs in Rust made me think Rust would be impacted next. Personally I think third party modules that are effectively universally used (axum, tokio, serde) should be integrated directly into the language.
Well I think the main important argument here is you shouldn't use the package management. A third party package manager would be even more concerning to use imo. I generally try to stay away from the typescript world so unfamiliar with nuances of NPM.
It usually takes some time for an updated dependency to actually get shipped to users in a release, by which time there's a good chance the attack has been noticed.
> Why do none of these hijacks embed runtime attacks? It seems like worming the build machines is the goal, rather than compromising downstream users.
Developer machines are quite juicy targets. They tend to have all sorts of credentials lying around, so it often isn't too difficult to escalate from that to compromising AWS/GCP/etc.
On top of that there's very little preventing a compromise. Developers are inherently expected to run untrusted code, and the usual Linux / MacOS laptop probably isn't even running any kind of anti-virus protection. Want to compromise the downstream app? Now you also need to pass Play Protect & friends.
> It seems like we should be building and testing everything in bubblewrap or some other sandbox going forward.
Yes, we really should. It is frankly a miracle that it has taken so long for fetch-time / install-time code execution to start blowing up in our faces. If we are spending so much effort on run-time isolation, why are we completely ignoring all those practices during development?
Why are developers "inherently expected to run untrusted code"? Running untrusted code on developer machines seems like a terrible idea, given that it will compromise everything they build or deploy and (as you mentioned) all their credentials.
because cloning repos and installing dependencies is what developers do daily. 90% of more don't care or even know that doing so they are risking getting hacked.
How do we know they don't? Who is going to read the source for all these micro packages?
You might assume we are able to detect all malicious behavior at runtime. But "stuxnet" and more recently, the xz compromise say otherwise. What if a not-so-popular crate deep in the dependency chain subtly introduced a LPE in it's code?
Jumping off the title and ignoring contents of post, as is customary:
look at this graph and guess which languages use each number of dependencies for their website: https://www.youtube.com/watch?v=E82ly38YEEQ&t=325s
I won't spoil the claim in the video about what the dependency number is correlated with, and I'm not even sure how true it is in general, but it's very interesting.
I’m disappointed crates.io doesn’t have a stricter bar for serving a crate that has newly acquired a proc macro or build.rs. That seems like a trivial mitigation.
Compromising the code that is then most likely run in a test instead of compromising a build script is just a very slight inconvenience for the attacker.
I'm not particularly fond of arbitrary build scripts either, but restricting them will not help the supply chain issue in a significant way.
Also there are several ways to control build.rs execution in the Cargo ecosystem, for example with cargo-deny.
Defaults matter. It’s nice you can set this up using a plugin to protect yourself, but that doesn’t protect the ecosystem, most of which doesn’t use cargo deny.
I also disagree a build scripts is a mild convenience. A build script always runs for anyone it’s a dependency for with full access and context and often has access to secrets in CI. A compromised runtime has more limited access and requires actual invocation of code paths (if you’re lying as a dependency that’s never executed, no exploit).
Of course Rust should have language-level support for capabilities so that just invoking a function doesn’t grant it access to arbitrary disk access. But that’s a much more difficult change than tweaking the defaults for cargo.
So? Runtime code requires actually executing the malicious code path which isn’t an immediate 100% hit rate for everyone that includes it in the dependency chain. For build.rs it’s a 100% compromise of everyone it’s in the dependency chain for. Additionally, at runtime you may not have access to secrets whereas at build time you most certainly do.
Why this still happens? Why after many previous supply-chain attacks maintainers of package repositories still allow anyone uploading packages and pushing updates without security audit?
Who is funding this security audit? Are folks supposed to volunteer their free time? It's a difficult coordination problem. The best folks have come up is to delay adopting new releases by a few days and hope your dependency is popular enough that a security firm audits it for you in that timespan. If you have enough money I suppose you can start employing llms to audit things for yourself.
Well, it depends on the language: language which "capabilities" (both for the source code and for the building* ) could in theory really reduce a lot of the burden to identify supply chain attacks.
*: some research language have/had capabilities which would make supply chain attack "obvious" but for build systems I don't know if this exist.
> Who is funding this security audit? Are folks supposed to volunteer their free time?
Same people who keep the whole rust project going, a lot of those are volunteers aren't they? Not mad to think they could do the same for core packages at least
> Same people who keep the whole rust project going, a lot of those are volunteers aren't they?
Sure, but from my understanding the Rust project is generally "bottom-up" in that volunteers generally work on what they want to rather than submit their time into a pool for some kind of higher-level management to direct.
It’s absolutely mad and extremely entitled to expect that a volunteer group of developers do an order of magnitude or more additional work for no additional pay or benefits to themselves.
Not really, if you're putting out something like programming languages and tooling, people expect them to work.
Especially because Rust devs brag so much about how it's soo superior to everything else, but then these amateur mishaps happen.
Rust isn't getting the exposure it deserves, I think, partly due to arrogance within the Rust community and a mental complex about "being better than everyone else" - that mentality never works
Languages like Rust have sources of income to be able to finance such audit.
You don't need to audit all the crap is being uploaded right now. Only really necessary and widely-used packages should be managed in a centralized way, so, auditing all of them isn't that huge task.
Mozilla, Google and a couple of others are publishing their audits through cargo vet. There are also additional audits done by individuals you can use through cargo crev. Overall the number of audited crates is in the thousands and you will find audits for most of the popular crates.
In the end it is your decision to use unaudited or refuse unaudited crates.
> In the end it is your decision to use unaudited or refuse unaudited crates.
It should be the default behavior of the package manager to allow downloading only audited/trusted packages. Forcing end-users of the language to be responsible for audit of all dependencies is impractical.
These very small dependencies that are then later causing issues either due to malicious nature or incompetence, have become pervasive in computing (for some reason). I think that these should be less of an issue now than ever. Outside of the largest, most critical dependencies, you really shouldn't be pulling in small libraries anymore. Just generate the code via AI. AI is not great at large scale programming I think, but its amazing at snippets of code. Something I ran into recently, I needed to use FFT2 on some matrices, and what I was using didn't have an existing solution. Converting some numpy fft2 tests to my target language, and having a full native implementation of fft2, and an accompanying test suite so it will behave exactly like numpy fft2. A few minutes and a few thousand lines of code later, I have a trusted implementation. Saves me an external dependency, some weird glue code, and an attack vector.
GitHub really needs something finer-grain then just pretending the repo never existed during these incidents. [1]
The bad package version has also just disappeared from crates.io [2] with no indication its been yanked. There's no security advisory there either [3] "No advisories found for this crate."
I feel crates.io was unprepared for a security incident like this since they're managing the response [4]
[1]: https://web.archive.org/web/20260820145918/https://github.co...
[2]: https://crates.io/crates/arrayref/versions
[3]: https://crates.io/crates/arrayref/security (I'd give an Wayback link but that's also broken https://web.archive.org/web/20260820150747/https://crates.io...)
[4]: https://github.com/rustsec/advisory-db/issues/3161#issuecomm...
> The bad package version has also just disappeared from crates.io with no indication its been yanked.
So, yanked crate versions do have an indication on crates.io. (Here's an example: [1]) The Rust blog post uses the word "deleted", and I'm guessing a bit here, but I think they mean that literally, and that the version here is deleted, not yanked. And I think that would be more appropriate: a yanked crate is still downloadable by cargo, if your lockfile is locked to it already; yanking only prevents lockfiles from newly automatically acquiring a lock on that version. That wouldn't be desirable in the case of a compromised crate: you don't want a download occurring at all. The tradeoff of "break those with locks on the crate" tips to being worth it.
That said, I agree with you, though: I think this state (if it is "deleted" and not yanked) should be plainly indicated on the crate's versions page. (Even better would be if it came with a link to, e.g., the blog post or a RUSTSEC so that you could find out why.) (& I think perhaps the docs for yank should point out whether or not it is appropriate in the "compromised crate" scenario, and if not, what to do instead.)
[1]: https://crates.io/crates/aes/versions
> That wouldn't be desirable in the case of a compromised crate: you don't want a download occurring at all.
Sure you do; you just don’t want the crate to be made available to people trying to make use of the crate as regular downstream consumers. You still want the crate available for download for analysis. You especially want to be able to deterministically reproduce the vulnerable version of your software that you already built.
I would posit that most package ecosystems should treat “contains an exploit” as its own package state; mostly ignore them; but still be aware of their existence.
In other words, if there’s any other non-exploited option under your version constraints, the dep should resolve to the nearest non-exploited version, even if older. But if it’s the only version, or if you have pinned the exploited version, then the packager should see it, but normally refuse to interact with it—i.e. refuse to lock to it if it’s the only version; or refuse to fetch it if it’s the locked version.
I say “normally” because you should be able to bypass this with a flag / env-var to the packager that basically means “I’m building this for forensic analysis, not for running.”
I agree but think the problem is slightly more general. Consider that a package could be redacted not just due to exploit but also (among other things) for legal reasons, and that such reasons could be region specific. So a redacted package might or might not be available to manually download and some other metadata is desirable so that "audit" can notify you that you were compromised.
A package redacted for legal reasons should still be able to be built too! Presumably one or the other side of the relevant suit will need the package as evidence!
I would suggest that such packages should simply require authenticated access + manual “write an email to ask to have your account be put on a whitelist”-driven authorization. Such a package would still be there (in its redacted state); but attempts to download it would be a 401 (if unauthenticated) or 403 (if not on the whitelist) error response, which the packaging tooling would need to know how to handle.
(In any case, this “redacted” package state and my “exploited” package state could be two variants of the same logic, both following the same approach to locking/resolution, both requiring the flag to indicate that you understand what you’re doing and aren’t just going to run the resulting program. The only difference is that once you pass that flag, an “exploited” package would then just work, for anyone; while a “redacted” package would still have one more gate standing (server-side AAA) in the way.
I think "deleted" is still always going to have to be a package state, realistically. A package could contain spam, abuse, copyright infringement and/or illegal content.
I don't mind malware being lumped in to that - the place to look up code for historical or analysis reasons ought to be version control, not crates.io
What about the place to look up packaged binary releases for historical or analysis reasons, esp. when the packaging system either builds such releases itself, or where it allows the submission of arbitrary build artifacts (where the build process that produced said artifact can be compromised in ways that have nothing to do with the source repo)?
> the place to look up code for historical or analysis reasons ought to be version control, not crates.io
The build infrastructure should make security incidents easier to respond to, not harder. I shouldn't be sent on a wild goose chase at a time when I'm potentially already dealing with a major incident. The tools need to "just work".
[3] is no longer true. They're definitely not unprepared for an incident like this. It's not the first time they've done it and they published an update to their process in Feb:
https://blog.rust-lang.org/2026/02/13/crates.io-malicious-cr...
Not having an advisory INSTANTLY available isn't a sign of a decaying org. Chill.
> They're definitely not unprepared for an incident like this.
We shouldn't need to resort to pasting `find` commands [1] into the shell from blog posts to tell if we're compromised.
`cargo audit` should be reporting if these packages have been downloaded. The "What you need to do" section of the blog should be run `cargo audit`
> Not having an advisory INSTANTLY available isn't a sign of a decaying org. Chill.
The GitHub issue was opened 8 hours ago. The cargo.io team also acknowledged it 8 hours ago. [2]
> [3] is no longer true.
Previously published versions shouldn't just disappear from the list. There's still nothing there to indicate a version was yanked.
[1]: https://blog.rust-lang.org/2026/08/20/supply-chain-attack-on...
[2]: https://rust-lang.zulipchat.com/#narrow/channel/318791-t-cra...
> We shouldn't need to resort to pasting `find` commands [1] into the shell from blog posts to tell if we're compromised.
I know you recommended `cargo audit`, but that is not nearly as universally installed as `find`. And as far as I know, cargo-audit has to be run within a project directory? Can it be run outside of the context of a project and scan the cargo registry cache?
I personally appreciated having the `find` command:
* It very clearly indicates where to look (my cargo registry cache)
* It very clearly indicates what files to look for (a list of wildcards)
* It's something that I can easily review and then copy/paste into my terminal
* I can very easily adapt it to my particular environment (perhaps into a `fd` invocation if I'm on Windows, or perhaps adapt it to scan all home directories on my system), or feed these file names into some other vulnerability scanner
I believe what's being suggested is that `cargo audit` should be a part of cargo rather than an add on.
You'd obviously already have cargo installed, which means you should be able to run a command against cargo to see if you are currently exposed to any security problems. Even if it's not the current "audit" add on.
As someone using bazel at work, the find script was much more amenable to me trying to audit our usage - we don't use cargo, so no cargo cache or cargo audit, but I can still figure out whether a particular package was downloaded.
If you're using a different tool then it falls to that tool to address the same problem. The point is that I shouldn't have to already know that a compromise has occurred and then go looking for it manually. The tooling should automatically start shouting at me very loudly that something happened with one of the packages I'm using.
It wasn't yanked, it was fully removed.
They did that because a yanked package can still be used. And in this case there weren't any downloads, so it is unlikely the malicious package was actually used.
Call it how you want, the point is that it should not disappear like this. Maybe yanked has the meaning you said, but in Ruby yanked has the meaning OP said and that is what we want.
How to call it then : Yanked hard vs yanked soft ?
Crates.io currently says:
> A new version of the arrayref crate was published with a direct dependency on proc-macro1, which would execute a malicious build script.
> This compromised version was published on 2026-08-20 and removed approximately 86 minutes later, with no evidence of actual usage.
I don't know what more you want. Do you want the malicious version to continue to be available?
> I don't know what more you want.
The version page [1] should show that for 86 minutes there was a version 0.3.10, it was malicious and was deleted with a link to the advisory. I get this takes time so even a generic "deleted" entry until they have time to link in the advisory would also be fine so we know something is happening.
Presumably even "deleted" crate versions still have some metadata left behind in the backend so this should be surfaced.
[1]: https://crates.io/crates/arrayref/versions
I have generated a report based on my own exposure to this attack today. The report was updated as the attack unfolded; https://acje.github.io/systems/arrayref_incident_record/
> GitHub really needs something finer-grain then just pretending the repo never existed during these incidents.
Google should read this too. They simply remove Android apps and Chrome extensions without a single word. No page explaining why they removed it, if i was at risk.
If anyone wants to create a revision to RFC 9110 :)
I think we should be taking a more “batteries included” approach to language and library design. The entire reason we’re in this mess is because we’ve decided it’s ok or maybe even preferable if stdlibs are rail thin, rendering base languages near-unusable.
I can very easily build a highly functional, pleasant to use Apple platform app with 5 or fewer top level dependencies. In many cases, I reach for between 0-2 total.
There’s no reason why this can’t be replicated elsewhere. The key is to make the programming language reasonably robust with at least 80% of common non-UI dev needs built in and put the remaining 20% and UI bits into a small family of well-supported, community-embraced, preferably first party libraries.
That would make it unnecessary to pull in foreign dependencies in the overwhelming majority of projects. What few do get pulled in becomes lightweight, easily verifiable syntactic sugar or libraries with purposes too niche to be worth targeting.
Of course this approach can go wrong too. You could easily end up with a monster like Boost, but that comes down to project administration keeping creep under control and proper modular design.
This is spot on. I am extremely uncomfortable with the large number of transitive dependencies that end up in pretty much any non-trivial rust application. No amount of memory safety will save us if a tiny, ubiquitous library that nobody scrutinizes because it’s nested eight layers down the dependency graph gets compromised.
I think Go apps tend to have better dependency hygiene because the language has a better standard library, which results in better culture around dependencies.
I also think it’s frustrating that both cargo and npm totally ignore the decades of prior art from Linux distributions that have figured out good ways to improve dependency management. We’d be in better shape if there was a community curated subset of known-good dependencies that are release-managed together that the broader ecosystem could build on, sort of like Ubuntu having “main” and “universe”.
I learned programming from C# which has an excellent standard library. Just recently started using Go which seems to have a good standard library. Been using python for years, can't say I ever had a complaint about the standard library.
Rust meanwhile seems to be following some "no standard library" philosophy.
I feel like there's an opportunity out there to become like an amoeba: fund and build a third party Rust standard library, absorb absolutely everyone who is so grateful to have a library, and then get so big absorb Rust itself.
I think Go ends up with fewer dependencies mostly because there is more friction in finding them. With rust (and npm) you just "cargo search xxx", then "cargo add xxx". Go makes you web search and poke around github looking for something. It's not onerous or anything, but just that extra step slows things down just a bit. C projects tend to have the least because it's even more annoying to add them and you have to create build commands that deal with slightly different distributions (and learn makefile+pkg-config or cmake, or autoconf).
Certainly having a larger stdlib helps Go projects keep their deps down, but I don't think it's the real reason.
Actually I think this is more of a culture thing. Or maybe "is also" a culture thing.
One of the core tenets of early Go was the maxim "a little copying is better than a little dependency".
Probably because of this stance, they didn't even HAVE a dependency-management solution for years
I strongly agree the fewer dependencies the better, on average.
Perhaps, but I think it depends on where the Go devs are coming from. In my experience the lack of proper dependency management in older versions of Go didn't really lessen dependencies, it just made teams have to deal with annoying $GOSRC issues. But then I worked at a place where the devs used PHP (w/composer) and node before Go was introduced.
Personally I came from a C background so I tended to use deps more sparingly.
In the end it's a bit of a balancing act—lots of dependencies is a larger opportunity for these kind of supply chain attacks to affect you. Copying stuff into your repo protects you from that, but it also makes it way more likely that you'll miss security fixes, unless you're actively looking for them. Re-implementing what you need is also viable but it slows down the dev process.
I agree. Java has the same ability to quickly add deps, and in some case they can be sprawling, but it's still perfectly possible to develop complex apps without a lot of deps coming in.
Culture has a lot to do with it.
> Certainly having a larger stdlib helps Go projects keep their deps down, but I don't think it's the real reason
I disagree: having a big stdlib that encompasses a huge chunk of common functionality (logging, a plethora of network server/client protocols, etc) means that for basic needs, the question of 3rd party libs never arise, whereas for other language, the question is not if you need a dep, but which one. Making gorilla easier to find won't undo the cultural reluctance of adding a 3rd party dep when using the (very pragmatic, IMO) stdlib server is adequate to the task.
Go has an official package search at https://pkg.go.dev/
There’s no poking, unless you are into that sort of thing
>There’s no reason why this can’t be replicated elsewhere. The key is to make the programming language reasonably robust with at least 80% of common non-UI dev needs built in and put the remaining 20% and UI bits into a small family of well-supported, community-embraced, preferably first party libraries.
The big reason is money & time. Maybe less so today if you are happy with an AI generated stdlib, but I don't think it's fair to compare Rust (built by Mozilla, who's primary product is an open source web browser) to Swift/AppKit (built by Apple, 4.6T marketcap at the time of writing).
When Go was released it had an amazing stdlib, but that was because Google was funding it.
There's probably to truth to this, but I would weight the design intention behind the language just as or more heavily than available resources.
Swift was intended from the start to not only replace Objective-C and all of its use cases, but also take on new use cases and bring a number of features from other languages that were newer and had different dominant paradigms.
Certainly having a juggernaut like Apple behind it has been instrumental in its success in achieving those goals. That said, a community-lead project with similar aims could probably achieve those aims as well, given enough time.
I think it's just rare for enthusiast-led projects to set out to do such things. My theory on why things tend to go that way is that the types of people to start programming language projects tend to heavily lean nuts-and-bolts and theory with a purist/idealist sort of mentality that's more concerned ideological purity than practical usability.
So to sum this all up, a community-run practical, multi-purpose, batteries-included language likely needs to at least partially designed and helmed by product engineer types so it doesn't end up anemic and stuck in an ideological rut.
Python is batteries included, but all the batteries have corroded.
Look at C++'s long in the tooth STL.
You don't want to marry a language to fast aging libraries you have to support for eternity. Better libraries always naturally emerge.
Rust's decision here is fine.
The only thing I'd like to see is the ability to programmatically limit transitive dependency count or depth in Cargo. I'd also like crates to specify whether they limit their own deps and whether they have panick-y behavior or not.
I think maybe the right answer is to have a standardized, curated group of libraries pegged at some sort of LTS release that only backports security fixes. However, someone would need to pay for creating and maintaining this -- and then you wonder where the money would come from?
someone tried doing this for rust, but the reaction was that it was vibe-coded/low quality.
https://github.com/rust-stdx/stdx
https://news.ycombinator.com/item?id=48571266
Note that there are baby versions of this that are uncontentious, for example
https://blessed.rs/crates
this is missing the LTS release. but it is a curated group of libraries that are relatively uncontentious to recommend.
That or potentially fast tracking merging popular libraries into the standard library. Or at least concepts from popular libraries. Basically all the "most downloaded" crates on the Crates.io front page should be candidates for merging into the standard library.
https://crates.io/
Good ideas are pulled into the standard library.
https://crates.io/crates/once_cell
https://doc.rust-lang.org/std/cell/struct.OnceCell.html
Not everything should be eligible for this treatment. Certainly not an HTTP library or something without extremely widespread applicability.
Nearly all of my projects used once_cell, so it's good to see that pulled in. I wouldn't want to see anyhow, thiserror, anything opinionated, or anything domain specific in the std. That's a weight you have to bear forever.
Remember how long Python 2 -> 3 took, and look at the ancient and awful stuff in Python 3's standard library. Rust is not the right language for this.
The fallacy is assuming that standard != unchangeable.
One can very simply....make breaking changes. Yes it will require some work to update but that's already the case when you upgrade library versions.
The dependency hell doesn't help anyone here, the downside is much worse - lots of bloat, overgeneric libraries and awful supplychain security.
I’m a big fan of how C# handles this. The standard library is versioned like any other library. Your program is explicit about the major version of the standard library you’re pulling in. And all major versions of the standard library continue to work.
For rust, this would mean something like adding std = 1.0.6 to your cargo.toml. The advantage is it means the standard library can deprecate and replace features. Upgrading std is an explicit step by the developer. Annoying, but in the age of LLMs it should be pretty easy.
The one big question I can’t answer is what happens if your dependencies use a different version of std? Are std v1 Strings compatible with std v2 strings? Oof.
Can we pick and choose the good without the bad? If we agree stability is a net-negative we could just say there will be a list of blessed trustworthy libraries with extra scrutiny and bureaucracy (make no mistake this is completely necessary to what people are asking for) to make sure they don't go rogue. But without a guarantee they will stick around and get updates forever.
Rust has a sizable and well featured standard library at this point. I think it would be absurd to claim that the base language is "near-unusable" if you are using it for system programming, which is its intended use case.
Huge standard libraries make sense for Java, Go, Apple platform etc. because they are developed by giant enterprises that can manage the overhead. Thinner modular systems are a more natural fit for open source development.
True. And the rust standard library is pretty good. But I really wish we had a good, fast, small futures executor in std. And accompanying async variants of File and Socket and so on.
Async rust is a jungle of weird compatibility questions. People treat async runtimes like sports teams. (I know I do). I wish it were more like Nodejs where async is just built in.
It’s being used for a wide variety of other purposes however, and so perhaps it’s time to adapt to that reality.
While overhead is a real concern, I think it’s often blown out of proportion. Once a language achieves a certain baseline of stability and isn’t in constant flux and the standard library matures, changes become infrequent and maintenance load is low. The work is heavily front-loaded.
I don't see the problem with boost? Isn't that a perfect example of your well supported, community embraced option? I certainly feel much safer pulling something in from boost via the official debian repos than I do pulling a random package with npm or cargo or etc.
Primarily I think the underlying concern has to do with the pathway for authoring code. When you have contributors whose submissions are gated with a rigid third party process and where that third party is the one responsible for curating the code (as opposed to the author also being the curator as well as the publisher) then you have the possibility to catch a lot of wrongdoing before it succeeds.
> I don't see the problem with boost?
There are many issues that I've seen taken up with Boost, but my personal peeve is how it can make building projects that incorporate it a pain, both because of its sheer size but also because it can sometimes be difficult to appease with its own dependencies.
that seems unrelated to the security problem to me
> I can very easily build a highly functional, pleasant to use Apple platform app with 5 or fewer top level dependencies. In many cases, I reach for between 0-2 total.
> There’s no reason why this can’t be replicated elsewhere.
It is replicated elsewhere and is not limited to Apple; I can do the same using a tech stack of perhaps 2x techs and very limited (manually added) deps. I'm thinking Lazarus, which comes with most things I'd need for most apps.
> I can very easily build a highly functional, pleasant to use Apple platform app with 5 or fewer top level dependencies. In many cases, I reach for between 0-2 total.
> There’s no reason why this can’t be replicated elsewhere. The key is to make the programming language reasonably robust with at least 80% of common non-UI dev needs built in and put the remaining 20% and UI bits into a small family of well-supported, community-embraced, preferably first party libraries.
I wholeheartedly agree. Most standard libraries out there (when I say most I easily mean of 99% languages that I know) have pitifully tiny implementations. They at most provide for string parsing, io, a bare networking layer (that's at best a wrapper around the POSIX syscalls), threading and that's it. If you want to do anything more complex, you either have to roll it yourself or use an assortment of the "blessed libs" for any given language, that may or may not work properly, and may or may not even work well with each other. I got pissed off by this at some point so I decided to go ahead and roll my own standard lib that would actually have everything I need to.. actually develop apps. Now I know most people probably aren't going to be using it (mainly because its functional API isn't what most C++ devs want) but for me it's actually been brilliant.
This was actually the main thing that put me off of Rust. I get the argument for a small std lib. I just also don’t agree it’s worth it. Go seems to handle having a batteries included standard lib just fine.
People confuse what they want.
They do not want a big stdlib. The downsides are real (the stdlib cannot make breaking changes), and there are no upsides (except, maybe, for faster compilation, since std comes precompiled).
They want more official crates (e.g. `regex` and `libc` are official crates, maintained by the Rust project). And the Rust project does not oppose to that, it just doesn't have the funding.
> there are no upsides (except, maybe, for faster compilation
There are other good reasons to have something in std. Another big one is providing types for cross-crate compatibility. Like if I want to pass a String from one crate to another, I’m glad that string is defined in std so there’s an obvious type we can both use in our APIs. C - for example - does not have this luxury.
It would be nice if we had a standard way to declare a type as serialisable. Right now lots of crates have a serde feature to do this with serde - but they need to put this behind a feature flag (for obvious reasons) and this doesn’t work with other serialisation libs.
The other big one is futures. There’s still no futures executor in std. Async crates have to decide if they want to tie themselves to a single executor (like Tokio) or be compatible. There is no async stream api in std. No async file api. And so on. It’s a compatibility nightmare.
Who said a standard library can't make breaking changes? That used to be a norm some time ago.
There are upsides, your program is smaller, better security because a random kid can't pwn your deps, quality and interoperability. What's not to love?
> Who said a standard library can't make breaking changes?
The std making a breaking change is the language making a breaking change. Most mainstream languages guarantee stability, certainly Rust.
> your program is smaller
How so? It doesn't matter if the code is in std or a crate.
> better security because a random kid can't pwn your deps, quality and interoperability
That's exactly what I said: you don't need bigger std, you just need more official crates.
>Most mainstream languages guarantee stability, certainly Rust.
That's their mistake, right there... When you're upgrading software, you presumably want a better version. You can't have something better without changing it. It's a logical contradiction.
>It doesn't matter if the code is in std or a crate. It very much does because you don't need 95% of the random stuff in the crates. Even for something like rand, you need an xorshift and that's roughly it. 25 lines of code, sorted. You can even write it as a copypaste "dependency" without much fuss in practically any language. That, versus importing a whole rand library with lots of different algorithms, deps on crypto, tests, OS random-based seeding, customisability, etc.
>you just need more official crates
I guess that's one solution but it would probably just be a better idea to split std in two SLAs, one is guaranteed for core stuff i.e. the status quo, one is YMMV.
> Who said a standard library can't make breaking changes?
I don't want my language creating work for me. "Batteries included" is a feature of higher level languages. Rust is going to live for a very long time, probably beyond all of our lifetimes, and it needs to endure fads without picking up baggage and liabilities.
Rust could have a mandate where certain crates are "blessed" as recommendations. This could come with additional review, oversight, support, documentation, etc.
The Cargo.toml spec should have the ability to limit direct and transitive dependency count and depth. (As well as other things, like "nopanic" annotations/guarantees.)
The Crates.io repo should grow namespaces so we don't typo squat popular packages. It's much harder to miss organization names changing than package names changing.
On the other hand, there are some bad Go standard libraries that are frozen in time.
Yeah, but I will take a not great library that works everywhere the compiler does, than be at the whims of which platforms are supported by 3rd party libraries.
I can use most of the clunky Python, Java, .NET and if it must be, Go, standard libraries, than hunting down for dependencies with platform tier support and such.
Add more and more to a standard library, and you're going to start losing your "works everywhere".
Naturally there is a balance, still I think Java, .NET, Python, Go, Smalltalk, manage quite well, while having subsets to slim down when needed for specific deployment scenarios like embedded devices.
Platform support is a non-issue for every library being discussed. This is an array construction macro, it has nothing to do with what operating system you're running on.
Just because some trees have a specific trait, it doesn't mean the whole forest is the same.
which, as painful as it may be, is ok. Better to have a safe functional stdlib library than a exposed external crate (which then asks the question; what's the replacement...)
Those aren't the only two choices.
We also have the classic example of PHP with numerous not safe stdlib ways to use mysql.
To me the answer is still to vendor your dependencies and don't be on the bleeding edge of updates unless you're willing to invest the time into validating them.
> We also have the classic example of PHP with numerous not safe stdlib ways to use mysql.
I think it goes without saying that emulating PHP is rarely a good decision.
Which isn't actually a problem. You can ignore the bad standard library and use something different.
The Odin programming language does this! It has also decided not to provide a package manager.
Go took the same approach and ended up having to implement a halfarsed one when everyone started implementing their own.
Has Go not had the most secure ecosystem?
What is your critique of their approach? Is it not the case that `go get` is the only one which doesn't even provide a way for the person downloading to run the downloaded code until it is actually executed by the consuming codebase? That seems pretty sound by comparison.
I guess you're saying, "they had to" meaning they should've seen the need and provided it? I'll say this to that (imagined) take; plenty of useful software was made without it, and they got the job done when they knew the absolute most about what a good solution would need. I totally understand being annoyed at the fact that this is the story of every evolution in Go, but I genuinely think they're picking good implementations when they decide on them.
> What is your critique of their approach?
It’s half arsed, brittle and far from user friendly.
> Is it not the case that `go get` is the only one which doesn't even provide a way for the person downloading to run the downloaded code until it is actually executed by the consuming codebase?
If you’ve added the package to your imported then odds are your next step is going to build it. Thus negating any benefit.
I think the real issue is malicious packages entering package ecosystems. Whether your package manage executes on downloads or not is moot because you’ve still got untrusted code sat in your project imports, just waiting to be accidentally executed.
> I guess you're saying, "they had to" meaning they should've seen the need and provided it? I'll say this to that (imagined) take; plenty of useful software was made without it, and they got the job done when they knew the absolute most about what a good solution would need. I totally understand being annoyed at the fact that this is the story of every evolution in Go.
I think you’re being too charitable here. I think Russ Cox just didn’t want a package manager because C doesn’t have one. But ended up relenting after everyone nagged the Go team for years afterwards.
And really what they built was the bare minimum to manage package version pinning and updates. But it has none of the visibility that central package repositories have. So how do you know if a Go package has been compromised when the only source of truth is the compromised origin?
Ultimately I think devs need to think about their dependencies and decide which ones get pinned and a serious review before pulling. If the thing has got binaries, it gets a serious review. If it does anything with cryptography, it gets a serious review... etc.
I don't think automatic updating is a good idea at all. It's just the honor-system, and trust is a security flaw.
Thanks for the thorough reply. I personally haven't experienced brittleness/unfriendliness, but I have only written around 10k lines of Go. Not exactly a power user, but I like to think I understand it.
> If you’ve added the package to your imported then odds are your next step is going to build it. Thus negating any benefit.
I actually think this is the benefit! I don't need to go to a website to see what changed, I can just have a look at the code. The best dependencies have a changelog. Ideally I can look at a diff.
> I think you’re being too charitable here. I think Russ Cox just didn’t want a package manager because C doesn’t have one. But ended up relenting after everyone nagged the Go team for years afterwards.
That's entirely possible.
> So how do you know if a Go package has been compromised when the only source of truth is the compromised origin?
Fair point, and probably worse in this future we're in now that NIST is drowning in CVEs and has turned away from some share of them.
Spoiler alert! JavaScript has no language provided package manager.
If Odin gets moderately successful someone will probably reinvent it.
NPM came along in 2010, Javascript was huge before that.
It was only when people wanted a common approach to shipping both in browser and "native" that this glitch happened. I believe a standard library and "batteries" would have abated it entirely or resulted in a slightly less-bad situation. I do think Cargo is a slightly less-bad situation in many ways, and that it can be done better.
Sure. Maybe threshold is bit higher than moderately. But unless your language tries to sabotage itself by making code artifacts uncomposable (a la C/C++ where best way to compose libraries is through shell commands) some package manager will be inevitable.
Batteries also don't help if dependencies don't replace them. Arrayref functionality has been part of Rust std lib for a while now.
>I think we should be taking a more “batteries included” approach
https://www.youtube.com/watch?v=GZOuz-SG7-g
Funny how you skipped "I should build my own batteries" and when straight to "increasing and centralizing the duties of your main gratis 'vendor'".
I don't like the "bring your own" approach because unless one happens to be an absolute tour de force 10x engineer unstoppable god of a programmer (which most of us, myself included, are not), whatever you build is never going to be as well-rounded, fleshed out, and complete as something built by a larger organization, especially when it comes to UI libraries (which are monstrous projects if done right, e.g. meeting accessibility requirements).
I'm happy to contribute to a larger effort but anything I can build on my own is going to be a thin, flimsy happy meal toy compared to something with the backing of a company or well organized FOSS project.
There's also the time factor. Something you build on your own will not have decades of development and polish behind it.
I agree that building your own is hard and something that can be done by exceptional engineers, but software is famously a winner-take-all industry, even if the optimal strategy for lower percentile programmers and median programmers to import generic modules, as long as building your own results in the best product (which it does, whether importing frontend or backend modules, if you hand code something, it's going to be the optimal strategy for building a winner product. Even programmers that are median or below average probably have an incentive to aim for being exceptional, because the best value of median and below average programmers is not building median or low value software, but having a shot at building exceptional software.
Also the challenge of building a generic module is much larger than building your own. A generic module needs to have flexibility for many different options and integrations, when you build your own you build just what you need and tightly integrate it with the product
> Also the challenge of building a generic module is much larger than building your own. A generic module needs to have flexibility for many different options and integrations, when you build your own you build just what you need and tightly integrate it with the product
True but I believe overestimated. Usually rather than building what's needed, what gets built is what is thought to be needed (often a substantially smaller subset), and then over time you end up building a markedly poor version of a generic module.
with LLMs it's becoming more common to just vibe up anything you need that might be missing. even if there is an available package you can pull in. the most secure option as well
Thread on the post from main rust blog: https://news.ycombinator.com/item?id=49372853
Direct post link: https://blog.rust-lang.org/2026/08/20/supply-chain-attack-on...
Initial report: https://github.com/rustsec/advisory-db/issues/3161
Other vendor posts:
* https://www.stepsecurity.io/blog/arrayref-rust-crate-supply-...
* https://research.jfrog.com/post/arrayref-proc-macro1-crates-...
* https://www.aikido.dev/blog/two-popular-rust-crates-arrayref...
Cargo desperately needs sandboxing for build.rs scripts. It’s been attempted before, but didn’t go very far¹.
¹ https://rust-lang.github.io/goals/2024h2/sandboxed-build-scr...
Sandboxing is a mitigation but a very limited one. The library itself can contain a malicious payload -- rust code most often ends up as native code executed on the host.
I think we may need to enter a world where there are fewer, heavily audited, libraries and dependency depth is limited. Allowing unvetted dependencies to be installed by default is not a good way forward. App stores have a similar problem and I think a lot of lessons have been learned there that can be applied.
That world can’t exist unfortunately, because the minute someone wants a feature that isn’t in your core libraries, you just create a new library and we’re back to now.
That's why languages need sandboxing at runtime as well
This is exactly what PMG is designed for ie. install/build time process level sandboxing. It currently doesn't support cargo, but I believe the challenges are same.
Here is my learning building PMG:
Sandboxing is good when the workload is predictable, and the goal of sandbox is to guard against exploitation of vulnerabilities, like sandbox protecting chrome tabs (renderers). But unfortunately build scripts are not predictable, at least not in npm/pypi world and I have seen build scripts doing weirdest of the things which is no different from malware. When popular packages do weird things, build breaks and users end up turning off the sandbox. This is a perpetual problem to deal with while building sandbox (or any least privilege solution) to protect unbounded workloads.
https://github.com/safedep/pmg
The "How PMG Works" section on Github does not actually explain how it works
build.rs by design can run absolutely anything. There tons of build.rs scripts that invoke a whole-ass C compiler toolchain to build and link C dependencies...
It isn't so much a question of sandboxing build.rs, as fundamentally changing the way that foreign dependencies are integrated into the rust toolchain (i.e. moving from a rust-centric system like Cargo to something more general like buck2)
Worth noting that the JVM ecosystem doesn't have this issue because there is no notion of installing dependencies, and the package managers are just downloaders with nothing else. No build.rs equivalent.
To solve the problem of native/C dependencies, they just bundle pre-compiled libraries as data files.
Rust could do the same thing.
But then that shifts the issue. You've now got an opaque binary blob being injected into programs. What if it is malicious?
It may or may not be a problem depending on whether it's actually used, which is still better than "gets you instantly the moment you compile the app". A lot of codepaths and even whole libraries aren't invoked just because a program depends on them.
It doesn't really matter anyway, because nobody is reading anything in their dependencies before it gets downloaded. Malware can also be hiding in plain sight in source code, as this attack and many others shows.
ABI stability is significantly different in a VM language to a native language, and reified generics necessarily require source compilation. Binary-only development for Rust would be exactly as 'sort of not really' as C++, for exactly the same reasons. (There is also no notion of 'installing' dependencies in Rust, and build-time code being malicious isn't much worse than runtime code being malicious.)
But people say build.rs is used mostly to compile C dependencies, in which case the lack of ABI stability of Rust is neither here nor there.
The vast and overwhelming majority of build scripts are building C code, so the other solution is to move away from integrating with C dependencies to native Rust dependencies, in which case adding friction to build scripts would be less noticeable.
Just denying write access outside the build directory and denying network access would go a long way and won't break pretty much any well-behaved build systems.
Any C library that's also packaged by debian supports being built under these conditions because it's required for everything except non-free packages: https://www.debian.org/doc/debian-policy/ch-source.html#main...
> Just denying write access outside the build directory and denying network access
As the link you posted mentions, you need a tiny bit more than that: you also need write access to the temporary directory (/tmp and similar). Many build tools temporarily store files there; for instance, unless things have changed since I last looked, if you don't use the -pipe argument the C compiler stores its temporary intermediate files (preprocessor output, assembler input) there.
One of rust's strengths is it's ability to interface relatively easily with existing c code without having to rewrite absolutely everything in rust. I don't think that is something we want to give up.
build.rs changes nothing about how easy it is to integrate with C. What does simplify: figuring out how to supply library you need at build time.
Which is the result of how bad dependency managment is outside (i.e. DLL-hell).
Pretty much all other use cases of build.rs can be sandboxed. Well, there is sqlx that wants to connect to database at expansion time to compile check-queries (yew).
sqlx at least has the (optional) offline mode, where you "cargo sqlx prepare" once (which wants access to a db) and then you can build in offline mode which typechecks your queries against local files.
Although I suppose that's still doing a lot of shenanigans at compile time. It could be sandboxed pretty well (theoretically). I'd hate to give it up completely though, getting a compile time error when SELECT query params or return values have type mismatches is extremely nice.
sqlx offline mode being opt-in instead of default is what bothers me. You know what else can validate that your queries return what you expect? Integration tests. Shoutout to sqlx for #[sqlx::test] though.
I also wish that it would be the default. Integration tests are fine, but they aren't compile time. Elevating them to compile time and using an LSP enabled editor makes it just underline SQL errors before I've even had a chance to run a manual compile let alone a test...
LSP diagnostics is actually what made me switch from compile time check queries. Whole "is db up? are migrations applied?" dance tired me pretty quickly. Its fine if you use sqlite, but anything else gets annoying.
Sandboxing for build scripts can't work properly. If you sandbox too much, some necessary stuff can't be done. If you sandbox too little, it has no practical value.
As an easy start, how about letting build scripts read /usr, read and write a temporary build directory, have some /tmp scratch space, and be allowed to write its final output artifact. No network and otherwise isolated from the rest of the system.
I would argue that, if a build script doesn’t work in the setting, then it doesn’t deserve to be installable by a default cargo command.
Cargo is a cross-platform tool, so when it ships a sandboxing solution it will need to be a cross-platform solution, and because this is a security feature it needs to be bulletproof, so no half-measures like Docker. Something like a WASM runtime might fit the bill, though that will be much easier to get working for typical proc macros than for typical build scripts. If you only care about Unix, then you can do this yourself today by building code in your sandbox of choice.
> Cargo is a cross-platform tool, so when it ships a sandboxing solution it will need to be a cross-platform solution
This seems like an excuse, not an actual objection.
Linux can do seccomp or Landlock or gVisor or a combination. Seccomp and gVisor need no privileges. Windows has its internal weird mechanisms. Mac has sandbox-exec.
Cargo could easily pick an appropriate sandbox for each major platform and ship it by default.
> If you only care about Unix, then you can do this yourself today by building code in your sandbox of choice.
This is ridiculous. The sandbox should not have network access, but cargo needs network access to download the package in the first place.
So let each build script define its own level of sandboxing and then users can determine whether they are okay with that level or not, e.g. `cargo build --sandbox-level=...`
That’s not solving the problem, that’s avoiding it by making it the users fault if they make a mistake.
Rust, like C, C++, and every other systems programming language, is all about giving users the power to make mistakes. The philosophical difference when it comes to Rust is simply that it tries to force the user to flip off the safety on the gun before letting you shoot yourself in the foot. A Cargo config option letting people opt-out of sandboxing would be fully in line with Rust's philosophy.
Sandboxing just build.rs would only be be a minor inconvenience for the attacker, nothing more. The attacker can always as easily compromise the binary you build and as soon as you run it (e.g. in a test) you are owned.
It would be a big pain for many that are in the unfortunate position to really need build scripts, though.
I imagine it would be sandboxed by default with an escape hatch to run build scripts outside of the sandbox with user verification. It makes people stop and think about what’s happening. Not perfect, but it does help. When working on JS ecosystem projects I manually approve build scripts and spend some time researching dependencies with build scripts to see if I can avoid running the build script. Some people will ignore it and run everything, but it’s a huge step in the right direction to make it operator-decided.
It would be more than a minor inconvenience. I can handle sandboxing my tests and production infra, but I can't handle sandboxing build scripts because I don't own that code in any sense.
At the very least, it wouldn't be overly onerous when adding a dependency that requires a build script to require an opt-in via Cargo.toml, e.g. `build-script = true`. You'd make it viral so that any transitive dependency that requires a build script would affect its parent, then add the key as defaulting to `true` so as to not break backwards-compatibility, then switch the default to be more restrictive over a new edition. (This same key could be used to prevent proc-macros from having arbitrary system access as well, where by default proc macros could be compiled to WASM and run in a WASM sandbox and treated as pure functions.)
I'd like to see a "no-build" option to blocks depending on crates using build.rs
I mean sure, but anything the build script could do, the build artifact could also do, That is to say, if you don't trust your source why do you trust the thing it compiles into?
Never going to work. Crates must be audited for behavior before use.
cargo add + rust-analyzer instantly executes build.rs before you have a chance to audit the code.
Cargo, please PLEASE give me a way to disable third-party build.rs and whitelist the ones I need. And please loudly mark any update that adds a build.rs where there was none before.
cargo-deny can audit build scripts, but unfortunately not prevent execution of malicious build scripts exactly for the reason you gave. It could still help if you only ever use cargo add and update in a sandbox.
See https://embarkstudios.github.io/cargo-deny/checks/bans/cfg.h...
You can disable rust-analyzer running build scripts.
So, don't add dependencies before you audit the code? That seems like a pretty reasonable ask to me.
Safe-ish recommendation: Only add well-known, trusted crates. Failing that, treat unknown crates as malware or containing malware dependencies until proven otherwise. Test untested tools/crates in a VM/dev container to be sure they work properly before trust them. Use Mark I eyeball and Mark III brain too. :)
Audit and limit crates (cargo-deny &| cargo-crev, && --offline) until tools exist to audit build.rs safety semi-automatically ($$ safeguard.sh maybe).
The root problem is two parts:
1. crates.io doesn't do mandatory curation. Lack of curation is fail. It's time-consuming and costly for reviewers without a doubt, but so is letting an ecosystem gain maximum entropy (go to shit) by Tragedy of the Commons depending entirely on the honor system. Name squatting, low-quality, unmaintained, typosquatting, and malware are the consequences of too much self-service / semi-self-service freedom.
2. Many, many cargo subcommands are over-eager to run build.rs because it assumes trusted crates:
In an untrusted/uncurated crates world and a build.rs exists or exists in a selected dependency, it shouldn't run at cargo-add time (unless it must). If it exists, on first run, it should be presented to the user in a viewer for manual review unless a magic CLI flag/env var is specified to accept it.
These 2 factors combined appear to create a Swiss cheese holes failure mode for running arbitrary crate `cargo add`. I have confidence a suitable add-in workflow &| standard command &| repository workflow will be adjusted to reduce the attack surface of the ecosystem.
I think it would be a good start if crates at least had to opt in to a build script, and adding one later would require permission from crates that depends on it.
The vast majority of crates don't need build scripts, so it is vaguely feasible to audit the list of crates you use that might need them.
https://news.ycombinator.com/item?id=49374811
(Oh and btw, proc macros also run arbitrary code.)
There's no good reason a proc macro can't run in a no-IO sandbox by default. None. Doesn't require a language change. Doesn't require some microvmcapabilityeffect BS. It requires looking people straight in the eye and saying "no" when they complain about needing to prompt for privileges.
How many times do we need to learn that sandboxing won't magically save us.
Rust suffers from the same faults as the JS ecosystem. Any significant crate imports hundreds if not thousands of dependencies. The probability that one of the authors gets targeted by AI-assisted attacks is just too high.
Also most of these dependencies provide a breadth of features that the end package does probably not need.
My experience has been that it has a major advantage, in that freeze + offline actually work properly. You can collect the dependencies you need once, put them in version control and never ever talk to remote registry again
A language without a large stdlib pushes this that functionality into (transitive) dependencies. I hope more language will adopt batteries included approach.
No stdlib will ever include all the "batteries" you need. And large stdlibs have their own set of problems, like stagnation, tying the library version to the language version, backwards compatibility garantees preventing evolution, and either needing a wide range of domain experts to maintain it, or having developers maintain components they don't have a deep understanding of.
I think the sweet spot is having an ecosystem of "blessed" libraries that are reviewed for security and quality, but are versioned independently of the language, and maintained by subject matter experts.
Yep. Sorry for the spam, but look at this, compiling "yazi" from source (ie a relatively simple TUI file manager) 676 dependencies:
.... Compiling clap v4.6.6 Compiling tokio-util v0.7.19 Building [===============> ] 451/676: ecdsa, exr, yazi-shared, regex-automata,'bs58' jumped out at me as a random looking package name.
It implements the base58 encoding, which is used primarily by... Bitcoin.
I'd love to see the explanation of why "a TUI file manager" needs a crate closely associated with crypto coins and not much else.
It is precisely this kind of thing that makes Rust a no-go for most enterprises.
"I just want a CLI too."
"Congratulations, your servers are mining crypto!"
Yes, the lack of a language package manager for older programming languages is a feature.
Doing software development outside of strict containerization, at the very least, looks increasingly prone to disaster.
Yes, we can argue about the culture of package management (as some of us have with especially npm from day one), but it's done, and your colleagues or AI sidekicks cannot be trusted not to download whatever and try to build and run it. All you can do is limit the effective blast radius.
Proper and easy to use support for sandboxes at the OS level, or better yet capabilities, seems like the only long term solution.
Many things I run I want to limit to r/w a single dir, and to have to request permission to make network calls.
Two problems with this:
1. There isn’t a single universal standard for sandboxing across all the different platforms that are supported by Rust.
2. Even if there were, if you’re compiling untrusted code then why would you trust the built output?
If you’re building the create then I’d argue that any preventative steps afterwards is akin to closing the barn door after the horse has already bolted.
> There isn’t a single universal standard for sandboxing across all the different platforms that are supported by Rust.
I said OS level. It's something I should easily be able to do via the OS capabilities that would work for rust, npm, etc.
I would use it not only for rust builds but for nearly every app on my computer.
> 1. There isn’t a single universal standard for sandboxing across all the different platforms that are supported by Rust.
Yeah, let's hold up the entire world of offensive cybersecurity capability while you work on that. Sure they will wait.
> 2. Even if there were, if you’re compiling untrusted code then why would you trust the built output?
You don't. You sandbox the hell out of it too.
minimum-release-age
Cargo has been working on min-publish-age, and the PR to stabilize the feature is in its final comment period and currently expected to land in Rust 1.100: https://github.com/rust-lang/cargo/pull/17335#issuecomment-5...
> rust-rfcbot commented 2 days ago: This is now entering its final comment period
I wonder if this is not in part the same situation as we arguably had with the xz compromise: some imminent change would have made the attack harder (in xz's case, IIRC a change to systemd to dlopen the compression library instead of directly depending on it), and the attacker rushed before the opportunity window closed.
is there a way to set up without using the nightly build?
what if I need a dependency my teammate released 5 minutes ago
You add company packages to an exclusion list.
It explains that in the link posted.
When several supply-chain attacks hit the npm ecosystem a few months ago, I built SBE: https://github.com/tyrchen/sbe.
It provides sandboxing for arbitrary CLI commands using Seatbelt / SBPL on macOS and Landlock LSM + seccomp-bpf on Linux. You can use it to protect local dependency builds, or integrate it into GitHub Actions to add an extra layer of protection to CI.
Feel free to give it a try — I’d love to hear your feedback.
We need effect based languages now. It's the only way to guarantee policies like no network, no file access, no unsafe code or FFI for a library before it even compiles. If anyone from epic is reading this please give us a timeline for open sorucing the Verse compiler.
In the mean time I think it's possible to hack Cargo and run all build scripts in a microVM. The blast radius will be limited to malicious code in the binary instead of uploading all your CI secrets and deleting the whole hard drive.
I think this is an attractive chimera, but will never be popular and will only be used much in specific domains.
The thing is, trust is a massive optimization. If you believe code from X is competently made and not malicious, you can accept that code without any additional overhead. (Building trust is a significant effort, but trust begets trust, so there's a resonable path from starting small to scaling.)
It's not that these mechanisms that require less trust (none of them eliminate it, BTW, when you examine them) can't work. It's that they'll inevitably be outcompeted by trust-based systems (in most domains).
E.g., compare the effort of creating a new, full featured, general purpose language, promoting it, building and maintaining all the tooling and documentation around it to support it, etc to the effort of creating and maintaining a curated package repository of trusted packages (and/or adding trusted code to extend the standard library) for existing language X.
(It's not clear to me you could provide enough value from a new language focused on effects that you could convince any but a few from adopting it at all, even if it dropped from the sky fully formed with all tools, docs, etc. in place from day one.)
> It's the only way to guarantee policies like no network, no file access, no unsafe code or FFI for a library before it even compiles.
It's not the only way. You can sandbox processes. I think sandboxing is the much more reasonable approach, because in the end of the day, there are still closed-source software products where you can't demand that the manufacturers use certain language safety features.
Sandboxing the process only works well when the malware requires more capabilities than the software itself.
So if your software needs to make HTTP requests and read the filesystem then the malware will be able to make HTTP requests and read the filesystem, which is enough for a ton of malware. Sure, maybe you can limit the directories it can access a bit and possibly some sort of network filtering, but it isn't a silver bullet.
Capability security in-language would make a huge difference, because the more granular you "sandbox" the less likely it is that the compromised component has the access it wants. For example if the HTTP client library is compromised maybe it can't access the filesystem so can't steal your cookies. Or maybe like in this case no permissions were needed at all and despite the process having enough capabilities for malware this malware can at worst return bad values and try to chain this to an exploit which is far more difficult than just running it itself.
The problem isn't the language here, it's cargo executing build.rs from dependencies which necessarily allows arbitrary code execution.
(as I and others have mentioned in the thread): The attacker can just move the malicious code from build.rs to lib.rs (ie. build-time -> test/execution-time).
Then the problem is the language, as the grandparent observes.
Running what you've built when you've added a malicious dependency also causes arbitrary code execution.
I don’t even understand what a capability based language is, but presumably closed source software wouldn’t apply here? Is it common for closed source software to give the source code to customers to build?
> Is it common for closed source software to give the source code to customers to build?
Yes, somewhat common. I work in embedded, and where a supplier gives us the ability to build their code it is a lot easier for us to deal with the next time we need a new build from them.
Closed source software could apply too, if the execution model supports it (e.g. WebAssembly). Just because the format is binary and "unreadable", its permissions (accessible functions) don't have to be
I think the only language that's used in production where malicious dependencies can't do arbitrary effects is https://roc-lang.org - but it's a pre-0.1.0 language and as such its production usage is extremely minimal for now.
Some shameless self promo - I'm working on one!
https://hale-lang.org/proof/
https://github.com/insanitybit/witchy
This is why I'm building this language. It's capabilities based. Build scripts can't just do whatever the hell they want to, everything is auditable, and it layers its sandboxing.
It's for fun and anyone looking should understand that this is AI driven building with human driven design, but the goal is to demonstrate.
I have been working on a cross platform way of sandboxing without a VM:
https://github.com/brianv0/formwork
Can you explain how an effect based language helps here? Also to note filesystem access can allow network access on many systems and also the other way around.
More than that, we need capability-based languages. No capability passed to it, no permission.
Why not object capability languages: https://blog.plan99.net/why-not-capability-languages-a8e6cbd...
No. Object-oriented brings things many languages cannot use. Also, that was a tl;dr fluff blog with barely any examples.
All languages in mainstream use are imperative and at least somewhat object oriented, other than SQL. But many of those don't have what it takes to implement capabilities (too flexible/dynamic).
First we need capability-based OSes, like we should have had decades ago if worse-is-better hadn't stuck us with Unix. I don't need to care whether or not a program was written in a capability-aware language if the OS fundamentally takes care of that for me.
HarmonyOS is a capability based OS. I don't know why anyone in China is still using Rust. The biggest social engineering hack was for the Rust team to convince developers that it is a safe language.
Actually HarmonyOS NEXT is getting its language as well, Cangjie.
https://cangjie-lang.cn/en
https://cangjie-lang.cn/en/docs?url=%2F1.0.0%2Fuser_manual%2...
FreeBSD does this with Capsicum: https://wiki.freebsd.org/Capsicum
1. seL4 exists. I haven't seen much more formal and rigorous than it. Efforts have ported the Linux kernel and RTOS-things to it.
2. Add syscall(s) to POSIX, which encourages *nix adoption, with an API like the following:
- Drop capabilities for thread/process except allow list.
- Test if current thread/process has all listed capabilities.
- List should be able to represent capabilities and their attributes, much like a single unveil() OpenBSD call, maybe as an array of/pointer to struct.
- (Program crashes if syscall capability is missing.)
Why does every build have the ability to run arbitrary code by default in the first place? That seems like something you should opt into only under very specific scenarios and even then it should probably be built around a WASM sandbox or microVM (as the parent suggested).
Because many many builds end up doing something just a little weird that the build system cannot handle by default. Ideally the build system would be fully property/dependency based and so it wouldn't have to run arbitrary code. In the real world everybody has something weird about their build that the build system cannot make work.
This is partially because for nearly every project the build system is something they need and don't care about. When they need something weird they hack just enough to make it work and never ask "how should the build system change so that this was a property instead of running code", and thus build systems are slow to improve. In a number of cases the build system did have a way to do that thing, but the person didn't know about it.
There are also a lot of code generators out there. I have yet to see a large project which didn't have their own code generator for something specific to their project (protobuf is an example from Google that has escaped and become useful elsewhere, but there are many others that are specific to one project. Yacc is from the 1970s, and stands for "yet another compiler compiler" - implying the idea was already common 50 years ago). You cannot have/use these useful tools without running arbitrary code.
I don't mind code execution as even make has shell scripting embedded, same thing with meson. My issue is with network access to download random stuff without it being declared somewhere and signed.
I'm using OpenBSD and the ports system set up two users `_pbuild` and `_pfetch` for building packages. The first one is for building and the sample `pf.conf` (firewall) forbid it from accessing the network. The second does fetch the files , but it's declarative with every files listed and signed. Even for languages like go and rust.
Fetching files while building is the bad idea there.
Good point. Many build systems also want to be a package manager, but the two are and should be separate. If you have a simple problem it makes things easy to combine them into one. However they need to be separate anyway, both for security and also to make other weird situations easier.
MicroVM this, effects that --- can we discuss security without needing to invoke bay area buzzwords?
The idea is "least privilege", and we desperately need it in computing. The precise technical mechanism we use to achieve it is less important than committing to the idea that a dependency doesn't run with full privilege of its host program and an install script doesn't run with full privilege of the programmer.
The original Linux seccomp is old enough to drink. It's always been possible to do things like expand macros in a no-IO environment. Nobody's bothered to do it over the past two decades. Why would anyone bother in the next two decades?
The technical mechanism is exactly the issue to figure out, there's a reason why projects don't have this and it's because different implementations have different tradeoffs.
No, the reason people haven't been doing this -- and we've had technologies for ages -- is that it's a huge pain in the ass, a "tax", that it's hard to get developers inside a company to pay, much less participants in open source ecosystems.
Look at how snap, flatpak, etc. provoke people to just turn off security rather than deal with breakages.
A new language won't help because the problem is social, not technical.
FFS, you can't even get people to use filesystems via intermediate objects in today's languages. People think your language is broken if you don't have an ambient open(). You don't need a new language to enforce capability discipline. You need to whack people repeatedly with a cluebat until their laziness and brain damage abate.
You can't simultaneously say that we've had the tech for years and also say that it doesn't work. We have not had the tech for years. x-plat sandboxing is extremely difficult, especially in a way that's performant.
Sandboxing almost always involves having a dedicated, privileged service, due to how operating systems have designed things. It requires platform specific code.
This is a technical problem and a social problem, there's zero reason to believe it's just one.
The technology works fine if you use it in a disciplined way. The problem is that people don't! You're necessarily going to break programs if you put them in restricted environments. That's not the same as the restriction technology not working properly.
> Sandboxing almost always involves having a dedicated, privileged service, due to how operating systems have designed things.
Or just invoking bwrap. Or sandbox-execute. There's no performance overhead. Complexity is minimal. Just read current Codex source code. It's not so bad. People have been making these sandbox tools for AI agents for years. Just need to apply sandboxing to all domains.
You want some kind of silver bullet that makes code "safe" without anyone having to change anything? To prompt for nothing? To use no IPC portal? That's not happening.
The people saying we need a new language or something are half right. We don't need a new language. A pure library solution is fine! But people do need to do work to make an ecosystem based on capabilities and least privilege to work. And we need to tell people who refuse to do this work to go to hell.
A new language is neither necessary nor sufficient. What we need is new set of balls.
> The technology works fine if you use it in a disciplined way.
Then obviously we have to discuss the implementation!
> Or just invoking bwrap. Or sandbox-execute. There's no performance overhead. Complexity is minimal. Just read current Codex source code. It's not so bad. People have been making these sandbox tools for AI agents for years. Just need to apply sandboxing to all domains.
Tools like bwrap literally could not have existed until very recently without also requiring suid/ privileges, and even today unprivileged user namespaces are not universally enabled. This is why the implementation matters. No, unprivileged sandboxing has not been around for years, especially not x-plat. bwrap is a perfect example of what I'm talking about, great reference - it either requires root or it requires unprivileged user namespaces and it's not x-plat. Great.
> You want some kind of silver bullet that makes code "safe" without anyone having to change anything? To prompt for nothing? To use no IPC portal? That's not happening.
I have no idea why you think I've said this, I'm pointing out that the implementation and technology matters deeply.
> The people saying we need a new language or something are half right. We don't need a new language. A pure library solution is fine! But people do need to do work to make an ecosystem based on capabilities and least privilege to work. And we need to tell people who refuse to do this work to go to hell.
Again, you can't say "this technology has existed for years! You can do it in a library!" and then say "but they don't!". You misunderstand the complexity and assert that it's merely a matter of will. It is both.
Browsers have been sandboxing their tabs since 2008. Don't sit there and tell me that bwrap, sandboxing, whatever is some kind of recent innovation with uneven distro support.
And yes, bwrap needs root or namespaces. So what? So does sudo. Setuid binaries are as old as time.
You know what's also been around and doesn't require root? Landlock. Have you heard of it? Has anyone? There's a huge PR and visibility gap in this space. We don't have to invent some new thing. We have to get people to learn about what we already have and use it.
"Oh, the technology works, but it need setup! Oh, my distro doesn't enable it! Oh, it doesn't work out of the box with my code!"
This is a social problem. It's not a technology problem. It's people not wanting to do the work, not understanding what already exists, dismissing solutions based on non-problems (like bwrap needing privileges on some systems), and in general adopting an attitude of "no", not a can-do problem solving stance.
And we're supposed to solve this problem with let another precious little effects language or another fucking MicroVM environment? FFS.
We need unprivileged sandboxing!"
"Use bwrap"
"Nooo, that needs setuid root or file caps!!!1!1"
"So the problem is that bwrap is part of the TCB?"
"Yes. We need sandboxing that relies only on unprivileged code like the Linux kernel"
... the fuck?
You can't solve a social problem with technology.
> Browsers have been sandboxing their tabs since 2008.
Yes, it's an amazing feat that has cost billions and led to major features like seccomp v2, ptrace sandboxing, etc. Do you know the history of browser sandboxing? It's pretty complex, a major technical feat.
> Don't sit there and tell me that bwrap, sandboxing, whatever is some kind of recent innovation with uneven distro support.
I can tell you that unprivileged sandboxing is not only new, it's ongoing and nascent work and not commonplace at all.
> And yes, bwrap needs root or namespaces. So what? So does sudo. Setuid binaries are as old as time.
These are massive footguns and issues for deploying code... You don't get that? Again, implementations have tradeoffs.
> You know what's also been around and doesn't require root? Landlock.
lol it is SO NEW what the fuck dude?
> Have you heard of it? Has anyone?
Yes, I use it!
> "Oh, the technology works, but it need setup! Oh, my distro doesn't enable it! Oh, it doesn't work out of the box with my code!"
> This is a social problem.
Yeah these sound like social problems if you have literally no idea what you're talking about lmfao. IT's BOTH.
Your ignorance is blatant and your position is dead in the water.
From the article: “[for Windows victims,] the [malicious] build script [fetches the attacker’s remote payload,] writes [it] to %TEMP%\rust-setup.ps1 and starts [it] through a VBScript launcher under wscript.exe, with a comment in the source explaining why:”
And the comment is:
So the malicious build script has a helpful comment (???), written in a familiar “terse nouns verbing” style (!!!).Would it be gauche to speculate? Maybe some script kiddy sweet-talked Fable into dropping its safeguards, or maybe Anthropic is doing a training run for Fable 5.1 and the air-gaps aren’t gapping.
Eh, just because the likes of Anthropic are only threatening you with their latest model, it doesn't mean older models can't do exploits...
All those folks telling me to update my dependencies, this is why I don't do it. It's not laziness, it's undeniable foresight.
As long as your versions don't have any security issues...
I mean, yes, update your dependencies. But probably after a week or so after they've been release and tested by the first penguins willing to jump into the ocean.
Holding periods for new updates are a good idea, but I would also prefer having a tool that could provide diffs of the entire project state pre- and post-update; including transitive dependencies being added or removed. Git diffs won't show you that information, by design.
What purpose does “undeniable” serve here? This tips over into hyperbole, in my opinion, whereas “it’s foresight” is much simpler and stronger. YMMV.
I think it was to make the tongue-in-cheek nature of the comment more apparent.
That’s undeniable.
There's a reasonable solution to this. Tools like `cargo-audit` can tell you if a crate has security issues and guide you towards an update. You don't have to blindly update all dependencies this way.
Nothing to do with Rust as a programming language, but meanwhile in the world of [language without a package manager], a monkey puppet glances awkwardly to the left.
The rust ecosystem is going to be hit by malware just like the NPM ecosystem. I have been saying it for years. They made the same mistakes or even worse mistakes because all it takes is a compromised serde to take the entire ecosystem down.
Here's a suggestion, do you think it could be this simple? If you write a package-manager, don't execute any of the downloaded code in an automated fashion. No hooks, no build-time metaprograms. If your language _needs_ metaprogramming to function, it's a huge secondary issue that I don't know how to solve. You can try to make a meta-program annotation which disables side-effects, but the metaprogram ultimately must write memory which expands the amount of code generated. All risk is introduced when the person downloading cannot preview the content in a safe place.
Unfortunately it doesn't solve the problem because malicious code can still end up compiled into the user's program which they will promptly execute, possibly in production...
Supply chain attacks are not a problem that can be solved by a single silver bullet, however the biggest benefit comes from a combination of minimum release age + fresh 2FA required for every publish + automated scanning. This makes it considerably more difficult to pull off a supply chain attack and should be the baseline security for all package managers.
The only safe solution is to review all the code's changes (or trust someone else to do it).
Forbidding compile/build-time shenaningans is trivially bypassable and has already been bypassed in the NPM ecosystem by just making the library code itself (not the build scripts) malicious - eg. do the bad thing when the code is loaded/tested, assuming the language has static constructors.
Post on the Rust blog: https://blog.rust-lang.org/2026/08/20/supply-chain-attack-on...
Discussion: https://news.ycombinator.com/item?id=49372853
> arrayref is a small crate of four macros.
Why do so many languages fall into this horrible practice?
There was a recent talk which explored this question (Dependency Cultures, by Richard Feldman):
https://www.youtube.com/watch?v=E82ly38YEEQ
Summary: it's cultural. Rust likely inherited the practice from Nodejs, who inherited it from Ruby. I think in Rust online spaces in particular there is also this undercurrent of "you're not smart enough to use certain parts of the language, so download libraries that handle that stuff for you."
The language also takes backwards compatibility very seriously, anything that goes into std "must be maintained forever". It is also argued that a large stdlib means the maintainers have less time to work on the language itself.
I understand not every language can have Go's amazing stdlib, but I would much prefer Pyhton's approach where every now and then some package/function from the stdlib gets deprecated/removed. Rust's 3rd party ecosystem is the worst thing from the language, worse than the compile times.
i'd like to welcome you to the hell that is c/c++ dependency management. Make? cmake? qmake? conf? autoconf? configure? autotools? submodules??? AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
Conan, Meson, Ninja, Bazel, meta-build tool / dependency manager #231 etc. etc. And then you have to deal with all the modern C++ and STL footguns.
It never ends .·°՞(っ-ᯅ-ς)՞°·.
I'd just select a tool, vendor the libraries I need to the codebase and call it a day. Did these for Eigen, liboption++, Catch2, Easylogging++ (now archived and not maintained anymore, sadly).
Build a simple makefile, and you're off to the races.
At least, for my cases, that is.
I'll take all that over downloading and immediately executing random unreviewed code any day.
None of those (other than submodules which is just vendoring) does dependency management. They are more configuration management than anything. C, Python, Perl (and maybe ruby?) relies on flag switches and environment variables to find all the necessary files and modules for compiling/running a script. Cmake and autoconf just configure those.
With NPM and Rust's focus on project's level dependencies, there's no longer emphasis on API stability. Instead we have breakage every months, forcing everyone on the upgrade treadmill. It's easier to audit C library because they focus mostly on security updates instead of redesigning the API for the nth time.
xmake, build2, whatever msvc and xcode insist upon…
vcpkg is pretty good. cmake is complex but very flexible. just need an llm to help and it's pretty much solved (if your dep is in autopkg)
The "npm-ness" of Cargo (centralized and standardized dependency management used at every opportunity) is generally pitched as one of the primary developer experience advantages over C++.
For those missing the npm like experience in C, they can use apt, yum, dnf, winget, nuget, conan, vcpkg.
There are two reasons you might want to use a crate, The first being that you want to use a good solid implementation that you know someone has spent more time doing and works better than almost any solution you could integrate. The second is you don't want to spend time implementing that.
Writing macros in rust is a pretty horrible experience but it's not difficult
just to clarify, by "horrible practice" you mean macros?
No, they probably mean tiny libraries.
Yes, that's what I meant.
Just like c++ thought threads wasn't a std library concern and then later changed they minds, rust will also change tac I predict
Rust is actively incorporating more functionality into the stdlib. The functionality of this crate has been in std since 2024. The ecosystem is just slow to update (not everything is maintained, etc).
The languages that have a poor standard library support have this issue and other languages encourage you to import tons of libraries to fix the problem.
This is why Javascript and Typescript suffer from this the most and has little to nothing to do with "popularity" and likely 9/10 of these npm packages import an external library.
Golang on the other-hand is just as popular and has a stronger standard library which people build against and it is encouraged to use its standard library rather than rolling your own or importing another package to solve the problem.
Rich official standard library vs "import tons of libraries" are not the only two options.
Java's standard library had arguably also been poor for a very long time and "import tons of libraries" just had not been practical for most of that time because the tooling and ecosystem for that did not exist yet.
The solution was apache-commons and guava. Two large libraries with everything the developers heart desired and well maintained by large organizations.
For Rust be probably will never have anything exactly like that because requirements from no-std development to fully fledged backend service are too diverse, but there is still room for a small number of well maintained backed by reputable developers convenience libraries in my opinion.
Namespacing. Mandatory namespacing. It's an essential prerequisite for any kind of reputation building and managememt, before you even get into the gritty technical or security details.
And it has to be mandatory. Top-level package names will always have more cachet. Developers are suckers for good package names, literal or imaginative. Plus it helps address, but by no means completely solves, name and typo squatting.
I understand people and groups can run their own crates.io-like repository, but that's a tangential aspect. Even if this were ubiquitous, you'd still want mandatory namespacing. You want provenance, or at least intended/nominal provenance, to be as transparent as possible, not implicit or buried. By no means a complete solution, but an important foundation for better technical and culture patterns.
I don't want to see a large standard library for Rust. If something is added to the standard library, then it is very difficult to change it afterwards because backwards compatibility.
It would be better to have blessed crates in crates.io. The Rust core team would release or audit them. If the blessed crates need breaking changes, it can be done by increasing their major semantic version number. That can't be done to the standard library.
Actually, there could be a "trust" level for crates: 1. blessed crates by the Rust core team, 2. trusted developers, 3. untrusted developers. Or something like that..
Go has a stronger standard library for certain use cases like basic CRUD web applications, but a lot of the Go standard library is also extremely low quality (flag, container, image, json, log, math, path, regexp, sync, time). Many of these aren't usable outside of toy use cases and have weird edge cases all over the place. Over time many will probably get new incompatible versions just like json. The container package in Go is the worst collection library in any mainstream programming language by a huge margin. It's actually astounding how bad it is.
FYI, json/v2 just dropped: https://go.dev/doc/go1.27#jsonv2 and log/slog has been a thing since 1.21: https://go.dev/doc/go1.21#slog
> a lot of the Go standard library is also extremely low quality
Now adjust your definition of "a lot" to include all other languages instead of apparently arbitrarily deciding that "a lot" means at best 10 and pretending that "weird edge cases all over the place" isn't normal.
I also wonder what your issue is with most of those, especially since e.g. regexp is excellent in the context of this thread - it runs in linear time, so it's not possible to craft a malicious input that will make it slow down to a crawl.
> Over time many will probably get new incompatible versions just like json.
What a bizarre statement to make.
FYI: "The encoding/json package is now backed by the v2 implementation.". You get all the benefits possible from the v2 implementation for free while maintating backwards compat and if you want to upgrade to something better, v2 is right there.
Do you have some magical suggestion how this could have possibly been handled better? And no, not having json in the stdlib isn't a viable path, that's just a cop out. But I guess since e.g. Rust and Java don't have json in their stdlibs at all, you can't say their json packages are bad, how clever and smart!
Another frequently made suggestion straight from la-la land is just writing perfect code from the get-go. Brilliant, can't believe nobody's ever thought of that.
Meanwhile in the real world, you've been able to reap the benefits of the solid encoding/json for over a decade now, and with v2 you get some nice free backwards compatible improvements and have a clear path to upgrading to v2. Perfectly handled IMO.
> Another frequently made suggestion straight from la-la land is just writing perfect code from the get-go.
The code in those packages is so far from perfect it's absurd. Like most Go developers, you just have extremely low standards.
I don't think this is true. People like to scapegoat the JS standard library but in reality most JS implementations have pretty featureful standard libraries, especially browsers. I've never felt the need to use any third-party libraries in the javascript projects I've worked on (except for maybe `ws`).
I don't think it's just that, though I agree they are clearly correlated.
The reason I don't think it's a sufficient explanation is that there is a clear history of large, 3rd party libraries being created exactly to supplement poor standard libraries. C++ has Boost, Java has Apache Commons (though Java also has a pretty huge standard library), arguably we could even say C has Posix/Win32/Cocoa.
I believe there is some deeper cultural reason why certain language ecosystems coalesce large utility libraries, while others prefer myriad tiny dependencies.
The important difference IMO is in the small vs big libraries culture.
Lots of languages have a bad stdlib but don’t fall into the trap of having thousands of micro libraries.
The reason people do it is because it brings clout and money. Just look for articles defending micro libs: the popular ones are by people who make a living on donations, due to maintaining 1000+ packages.
And collaborating in larger libs/stdlib is hard. Plus: Rust, Node, all have a lot of visibility.
You need a good stdlib culture to avoid it (like Go did).
The reason is simple: the package manager works well. Helps if the package manager is standardized and there is a "default" registry for open source projects.
I remember the days where I had to manually put the Spring .jar files into my project. No way I am doing that for 100s of dependencies.
I've never seen a non trivial Python or Go project without external dependencies. The dependency tree of comparable Go and Rust projects seem comparable, IMO.
Pytorch has 9 transitive dependencies for cpu-only execution, or 29 to bring in cuda.
candle, the most popular rust ml library I found in a cursory search, has 119 for cpu only, and 150 to bring in CUDA.
I guess it's taste whether that's comparable
this would have been a way more satisfying dunk if nvidia hadn't split the cuda functionality needed by pytorch into 19 (!) packages on pypi but such is life.
I expected somebody to pull out an example "disproving" my point. It wouldn't be hard. I can point out rust projects that have a ton of external dependencies, and then point at similar projects with very few.
But a Python example doesn't really count, in my mind -- Python is pre-GitHub so tends to have small numbers of large external dependencies, like C++ and other older languages.
If a python example doesn't count, why were your two example languages python and go?
But yeah, small number of large dependencies is the way to go.
Total nonsense. Python has a massive stdlib and there are malicious packages.
Can’t agree more. If you take security seriously, you should consider using Go.
Only the ones that make it easy. Language package managers are a mistake.
Unfortunely we haven't yet gotten a universal OS agnostic package, other than zip and tarballs.
They aren't trivial, at least: https://docs.rs/arrayref/0.3.9/src/arrayref/lib.rs.html#202-...
We really need Cargo RFC #3923 https://github.com/rust-lang/rfcs/pull/3923
It would allow projects to avoid newly published dependency versions until a configurable waiting period has elapsed.
npm has had a similar min-release-age feature for years.
It's a pity that progress has been so slow.
https://github.com/rust-lang/cargo/issues/17009
Unfortunately Cargo doesn’t have security controls in place to prevent these kinds of attacks. For example pnpm has controls to allowlist install scripts for dependencies and will warn about new install scripts (without executing them).
There is an open issue for this: https://github.com/rust-lang/cargo/issues/13681
Compromising the code that is then most likely run in a test instead of compromising a build script is just a very slight inconvenience for the attacker.
I share the dislike for arbitrary build scripts but restricting them will not help the supply chain issue in a significant way.
Also there are several ways to control build.rs execution in the Cargo ecosystem as well, for example with cargo-deny.
You’re right about attackers being able to change runtime code.
pnpm does have some other features to prevent supply chain attacks, so there is still something to learn from other ecosystems. For example pnpm has a cooldown period for new dependencies and can prevent trust policy downgrades (eg new version published without build provenance where older versions did have it). See https://pnpm.io/supply-chain-security
> pnpm has a cooldown period for new dependencies
Cargo has `min-publish-age` in nightly, and it's currently heading towards stabilization: https://github.com/rust-lang/cargo/pull/17335
Disallow lists are ineffective - better to disallow by default and require opt in.
Also, crates.io can defer serving up newly uploaded scripts that have a new build.rs / proc-macro dependency and warn publicly that a version introduces it.
Restricting build scripts 100% will help mitigate the impact, just not if you only deny it once. And they can develop other things like sandboxing for build scripts by default and escaping that to be the exception that has to be explicitly allowed.
The problem is that build scripts run automatically without user consent or intevention.
`cargo add` is sufficient to compromise you, before you have a chance to even vet the code.
`cargo add` just modifies your Cargo.toml, it doesn't build anything.
Are there any plans to more seriously develop the standard library in Rust? Or is the plan to remain in this status quo where users of Rust import nonsense and the dependency tree explodes (or users are forced to invent their own wheel?).
Are there any comparisons between the state of the stdlib in C++ vs. Rust? I’d think that would serve as an excellent jumping off point to start chipping away.
The standard library is not versioned, so any API*/behaviour there must be maintained forever. When you put it in a separate crate, you can make better interfaces or be displaced by a better crate, while old code still compiles against the version it was written for. So even code effectively maintained within rust teams (hashbrown, rand, regex) may be in separate crates, std is specifically for OS abstraction and a shared type vocabulary.
That doesn't mean you jump to importing nonsense or trivial dependencies. The top hundred are efficient, well-designed crates of the quality you'd expect in a large std like Go or Swift, sometimes even higher as people can write better implementations that they otherwise wouldn't (or wouldn't be used over std). And those languages make breaking changes! C++ is mostly stable, so it's full of junk like <regex> or just unordered_map. Rust managed to wholesale reimplement HashMap 6 months after SwissTable released, like it is also easier to express this level of encapsulation, but that's part of countless design/interface decisions made deliberately to not constrain forwards compatibility, including a smaller std.
However, that's no excuse to have a worse developer experience in this area. We need better tools to vet and communicate the quality of a crate and its supply-chain, like community-curated or even additional org-maintained crates, and maybe a handful delivered precompiled in the default rustup distribution which can change over time. I don't think they need to be added to std itself though
The idea that Rust has a small standard library is a weird meme. Rust has an enormous standard library. Look at any Rust release and you'll see dozens of new library APIs added, and consider that Rust has about nine releases a year, meaning that Rust adds over a hundred APIs per year, and has consistently for the past ten years. Rust frequently adds things that obviate popular crates, most recently the cfg_if crate was made redundant by the new cfg_select macro. And half of every dependency graph that people wrong their hands over are actually first-party external crates provided by either the Rust organization itself or by known contributors to the project. When people say that Rust has a small standard library, mostly they seem to just mean that it doesn't include a webserver.
It still lacks basic functionality like a JSON parser, regex, directory walker, rnd generator and cli arg parsing.
I won't mention lack of date/time lib because that's complex and changes often.
That's why projects end up with 100s of crates, sometimes 1000s.
This might not be a well received fact in Rust community, but it's a fact nonetheless.
My personal opinion on each of those:
JSON: there are a ton of different ways to do serialization and deserialization, each with their own tradeoffs, and serde (the most popular) is far from universally agreed upon. The same goes for JSON specifically, there are many different serialization formats with different tradeoffs.
regex: Owned by the rust-lang organization already. You get the benefits of trust (if you trust std, you trust the rust-lang organization anyway), but without the issues of being in std (backwards compatibility and bloat). The only problem I see with that is that BurntSushi is still the owner of the package and as such can still publish new versions on his own (AFAIK crates.io currently requires at least one user owner, but that's something that can be solved by improving crates.io permissions).
walkdir: It has been been postponed, due to the complexity of WalkDir, but may be added in the future. I agree this should be in std. https://web.archive.org/web/20260820171531/https://github.co...
RNG: rand is still evolving, with breaking changes half a year ago. Preferred generators tend to change over time, so I don't see those getting into std (remember, it has to be maintained forever!). I could see the interface (traits) getting into std, but I see no advantage to it: it's maintained by rust-random, but even if you wanted it to be maintained by the same authors as the Rust language (let's say you trust them more than rust-random), you could just move it back to rust-lang-nursery or rust-lang.
CLI arg parsing: not simple at all! The community's preferred solution has 70kLOC (with support for so many features), but there's also argh, pico-args, and others, each with their own tradeoffs.
> That's why projects end up with 100s of crates, sometimes 1000s.
Also because libraries are split up in many crates. For example, regex is split in 3 crates: regex, regex-syntax and regex-automata, and depends on other crates from the same author, such as aho-corasick.
All that said, there are many crates, such as algorithms like aho-corasick, that I think could me moved into the rust-lang org, like regex was.
See also: https://home.expurple.me/posts/a-big-standard-library-is-ove...
Rust seems barely better than Node in this regard. Go or .Net or anything with a robust standard library seems like the way to go for most projects.
Cargo (and PyPI) is undeniably better than NPM, which is just shockingly bad for cultural reasons. Yet it's not safe, and it's subject to the same class of exploit, as we're seeing.
Indeed, the solution is to get away from the wild soup of author-managed dependencies and go with something with an audited collection of software that is maintained by separate human beings from the known-vulnerable hackers writing the software.
And indeed Go and .NET and Java all qualify. But the gold standard here is Debian and all its downstreams.
> the solution is to get away from the wild soup of author-managed dependencies and go with something with an audited collection of software that is maintained by separate human beings from the known-vulnerable hackers writing the software.
I think we might be able to crowdsource audits. At least in the Rust ecosystem I'm confident that this is feasible with the right tooling.
Cargo-crev already exists if you care to use it.
It does, but the UX is pretty lacking compared to the regular package workflow.
Oh, so it's not only "JavaScript bad and npm bad". Apparently, if your language uses third party dependency registry, you are prone to malicious code, regardless if it's Javascript or not.
Use containers for development. And reduce the amount of third party deps you import into your projects. This is only going to get worse.
Pulling in lots of dependencies creates this kind of risk regardless of the ecosystem. That being said in the JS/NPM world you tend to have a LOT more dependencies (especially indirect ones) than other languages. I saw someone do a cursory analysis and JS/Node projects tend to have 5x the number rust or ruby projects.
This is really a cultural problem not a technical one.
It's funny how left-pad was always brought up as a JS supply chain vuln when that wasn't even malware
If it weren't a registry it would be ./configure scripts and makefiles. The issue is that sandboxing technology is kinda shit (especially x-plat) and languages don't build it in by default.
> The genuine arrayref and append-only-vec crates are maintained by droundy, whose account appears to have been compromised.
That name looked familiar to me; I believe it's the same David Roundy who was an academic at Reed College who wrote DARCS, which is version control software. I used DARCS when I started grad school around 2010 before switching to git.
What does the malicious code actually do?
I would like to know too, but the author's Github account seems to have been deleted, and the crates.io releases have been completely deleted too (not just yanked) so it seems impossible to view the file :/
Yeah I scanned the article and I couldn't find this either.
Rust finally got hit... This was motivating me to swap away from Rust towards because of the huge number of dependencies. It seemed inevitable. Ginger bill was right, https://www.gingerbill.org/article/2025/09/08/package-manage...
The author proposes to reinvent the wheel and depend on stale dependencies, which are EOL. Neither of that is an acceptable solution with LLM-based Agents being able to produce exploit(ExploitGym) chains in minutes from known bugs. The other issue is that Rust's forces the user to provide more information and APIs are usually kept generic for systems programmer, so standardizing things is not as straight-forward compared to Go where you can assume a memory-management, a virtual thread runtime and mostly ignore dynamic dispatching.
I don't think anyone is saying that you should depend on stale dependencies. You should not hand off your responsiblity to untrustworthy package managements services though. Evidence of the failure of automated package management is the near constant shai-hulud attacks as well as the originating post here.
The fact is that your dependencies are your responsibility. You should be staying up to date on what you depend on and their vulnerabilities. Most vulnerabilities affect specific code paths which may not be present in your code base and do not require mitigation. These are decisions you can make by understanding what your dependencies are instead of offloading to automated package management systems.
NPM has been the posterchild of these types of issues but the dependency stack I saw everytime I compiled even simple programs in Rust made me think Rust would be impacted next. Personally I think third party modules that are effectively universally used (axum, tokio, serde) should be integrated directly into the language.
Honestly, Ginger Bill is plain wrong.
Just because you don't develop a package manager for your language, doesn't mean someone else won't. See NPM.
Well I think the main important argument here is you shouldn't use the package management. A third party package manager would be even more concerning to use imo. I generally try to stay away from the typescript world so unfamiliar with nuances of NPM.
yep which people have already done in odin
I am surprised this took so long.
Why do none of these hijacks embed runtime attacks? It seems like worming the build machines is the goal, rather than compromising downstream users.
It seems like we should be building and testing everything in bubblewrap or some other sandbox going forward.
It usually takes some time for an updated dependency to actually get shipped to users in a release, by which time there's a good chance the attack has been noticed.
> Why do none of these hijacks embed runtime attacks? It seems like worming the build machines is the goal, rather than compromising downstream users.
Developer machines are quite juicy targets. They tend to have all sorts of credentials lying around, so it often isn't too difficult to escalate from that to compromising AWS/GCP/etc.
On top of that there's very little preventing a compromise. Developers are inherently expected to run untrusted code, and the usual Linux / MacOS laptop probably isn't even running any kind of anti-virus protection. Want to compromise the downstream app? Now you also need to pass Play Protect & friends.
> It seems like we should be building and testing everything in bubblewrap or some other sandbox going forward.
Yes, we really should. It is frankly a miracle that it has taken so long for fetch-time / install-time code execution to start blowing up in our faces. If we are spending so much effort on run-time isolation, why are we completely ignoring all those practices during development?
Why are developers "inherently expected to run untrusted code"? Running untrusted code on developer machines seems like a terrible idea, given that it will compromise everything they build or deploy and (as you mentioned) all their credentials.
because cloning repos and installing dependencies is what developers do daily. 90% of more don't care or even know that doing so they are risking getting hacked.
How do we know they don't? Who is going to read the source for all these micro packages?
You might assume we are able to detect all malicious behavior at runtime. But "stuxnet" and more recently, the xz compromise say otherwise. What if a not-so-popular crate deep in the dependency chain subtly introduced a LPE in it's code?
At least it’s safe
Jumping off the title and ignoring contents of post, as is customary: look at this graph and guess which languages use each number of dependencies for their website: https://www.youtube.com/watch?v=E82ly38YEEQ&t=325s
I won't spoil the claim in the video about what the dependency number is correlated with, and I'm not even sure how true it is in general, but it's very interesting.
I’m disappointed crates.io doesn’t have a stricter bar for serving a crate that has newly acquired a proc macro or build.rs. That seems like a trivial mitigation.
Mitigation for what?
Compromising the code that is then most likely run in a test instead of compromising a build script is just a very slight inconvenience for the attacker.
I'm not particularly fond of arbitrary build scripts either, but restricting them will not help the supply chain issue in a significant way.
Also there are several ways to control build.rs execution in the Cargo ecosystem, for example with cargo-deny.
Defaults matter. It’s nice you can set this up using a plugin to protect yourself, but that doesn’t protect the ecosystem, most of which doesn’t use cargo deny.
I also disagree a build scripts is a mild convenience. A build script always runs for anyone it’s a dependency for with full access and context and often has access to secrets in CI. A compromised runtime has more limited access and requires actual invocation of code paths (if you’re lying as a dependency that’s never executed, no exploit).
Of course Rust should have language-level support for capabilities so that just invoking a function doesn’t grant it access to arbitrary disk access. But that’s a much more difficult change than tweaking the defaults for cargo.
As mentioned by others it’s just as easy for an attacker to modify a crate’s runtime code.
So? Runtime code requires actually executing the malicious code path which isn’t an immediate 100% hit rate for everyone that includes it in the dependency chain. For build.rs it’s a 100% compromise of everyone it’s in the dependency chain for. Additionally, at runtime you may not have access to secrets whereas at build time you most certainly do.
The malicious code could use life-before-main hacks to gain code execution if it's linked at all, even if uncalled. https://grack.com/blog/2026/06/11/life-before-main/
It doesn’t take away from the point that build time often has access to secrets the runtime does not.
This is quite the egg on the face, given that Rust proponents keep telling us how it's great for writing secure software.
wow, what did payload do?
I hate cargo and npm. Why do we keep settling on arbitrary code execution in our build process?
ahh... we now have nodejs ecosystem attack techniques migrating to other systems as well...
It's not very surprising as the node attacks were very effective at gathering credentials.
damn!
I got to watch this attack unfold pretty much in real time as my agents worked the issue. Here is my writeup https://acje.github.io/systems/watching_a_supply_chain_attac...
Why this still happens? Why after many previous supply-chain attacks maintainers of package repositories still allow anyone uploading packages and pushing updates without security audit?
Who is funding this security audit? Are folks supposed to volunteer their free time? It's a difficult coordination problem. The best folks have come up is to delay adopting new releases by a few days and hope your dependency is popular enough that a security firm audits it for you in that timespan. If you have enough money I suppose you can start employing llms to audit things for yourself.
> It's a difficult coordination problem.
Well, it depends on the language: language which "capabilities" (both for the source code and for the building* ) could in theory really reduce a lot of the burden to identify supply chain attacks.
*: some research language have/had capabilities which would make supply chain attack "obvious" but for build systems I don't know if this exist.
> Who is funding this security audit? Are folks supposed to volunteer their free time?
Same people who keep the whole rust project going, a lot of those are volunteers aren't they? Not mad to think they could do the same for core packages at least
> Same people who keep the whole rust project going, a lot of those are volunteers aren't they?
Sure, but from my understanding the Rust project is generally "bottom-up" in that volunteers generally work on what they want to rather than submit their time into a pool for some kind of higher-level management to direct.
It’s absolutely mad and extremely entitled to expect that a volunteer group of developers do an order of magnitude or more additional work for no additional pay or benefits to themselves.
Not really, if you're putting out something like programming languages and tooling, people expect them to work.
Especially because Rust devs brag so much about how it's soo superior to everything else, but then these amateur mishaps happen.
Rust isn't getting the exposure it deserves, I think, partly due to arrogance within the Rust community and a mental complex about "being better than everyone else" - that mentality never works
The core packages (things like rand and regex) are pretty closely audited in practice (albeit it might not catch a credential compromise).
This crate isn't one of them.
> This crate isn't one of them.
still caught in hours though, so just as a general rule: never install anything newer than 7 days old packages
cargo feature for this is still unstable infuriatingly:
https://github.com/rust-lang/cargo/issues/17009
Languages like Rust have sources of income to be able to finance such audit.
You don't need to audit all the crap is being uploaded right now. Only really necessary and widely-used packages should be managed in a centralized way, so, auditing all of them isn't that huge task.
Mozilla, Google and a couple of others are publishing their audits through cargo vet. There are also additional audits done by individuals you can use through cargo crev. Overall the number of audited crates is in the thousands and you will find audits for most of the popular crates.
In the end it is your decision to use unaudited or refuse unaudited crates.
> In the end it is your decision to use unaudited or refuse unaudited crates.
It should be the default behavior of the package manager to allow downloading only audited/trusted packages. Forcing end-users of the language to be responsible for audit of all dependencies is impractical.
If you want that you can always use vetted package repositories like Nexus. Many companies do.
These very small dependencies that are then later causing issues either due to malicious nature or incompetence, have become pervasive in computing (for some reason). I think that these should be less of an issue now than ever. Outside of the largest, most critical dependencies, you really shouldn't be pulling in small libraries anymore. Just generate the code via AI. AI is not great at large scale programming I think, but its amazing at snippets of code. Something I ran into recently, I needed to use FFT2 on some matrices, and what I was using didn't have an existing solution. Converting some numpy fft2 tests to my target language, and having a full native implementation of fft2, and an accompanying test suite so it will behave exactly like numpy fft2. A few minutes and a few thousand lines of code later, I have a trusted implementation. Saves me an external dependency, some weird glue code, and an attack vector.