> This was the most critical vulnerability we discovered in OpenBSD with Mythos Preview after a thousand runs through our scaffold. Across a thousand runs through our scaffold, the total cost was under $20,000 and found several dozen more findings. While the specific run that found the bug above cost under $50, that number only makes sense with full hindsight. Like any search process, we can't know in advance which run will succeed.
Mythos scoured the entire continent for gold and found some. For these small models, the authors pointed at a particular acre of land and said "any gold there? eh? eh?" while waggling their eyebrows suggestively.
For a true apples-to-apples comparison, let's see it sweep the entire FreeBSD codebase. I hypothesize it will find the exploit, but it will also turn up so much irrelevant nonsense that it won't matter.
Wasn't the scaffolding for the Mythos run basically a line of bash that loops through every file of the codebase and prompts the model to find vulnerabilities in it? That sounds pretty close to "any gold there?" to me, only automated.
Have Anthropic actually said anything about the amount of false positives Mythos turned up?
FWIW, I saw some talk on Xitter (so grain of salt) about people replicating their result with other (public) SotA models, but each turned up only a subset of the ones Mythos found. I'd say that sounds plausible from the perspective of Mythos being an incremental (though an unusually large increment perhaps) improvement over previous models, but one that also brings with it a correspondingly significant increase in complexity.
So the angle they choose to use for presenting it and the subsequent buzz is at least part hype -- saying "it's too powerful to release publicly" sounds a lot cooler than "it costs $20000 to run over your codebase, so we're going to offer this directly to enterprise customers (and a few token open source projects for marketing)". Keep in mind that the examples in Nicholas Carlini's presentation were using Opus, so security is clearly something they've been working on for a while (as they should, because it's a huge risk). They didn't just suddenly find themselves having accidentally created a super hacker.
> Wasn't the scaffolding for the Mythos run basically a line of bash that loops through every file of the codebase and prompts the model to find vulnerabilities in it? That sounds pretty close to "any gold there?" to me, only automated.
But the entire value is that it can be automated. If you try to automate a small model to look for vulnerabilities over 10,000 files, it's going to say there are 9,500 vulns. Or none. Both are worthless without human intervention.
I definitely breathed a sigh of relief when I read it was $20,000 to find these vulnerabilities with Mythos. But I also don't think it's hype. $20,000 is, optimistically, a tenth the price of a security researcher, and that shift does change the calculus of how we should think about security vulnerabilities.
> But the entire value is that it can be automated. If you try to automate a small model to look for vulnerabilities over 10,000 files, it's going to say there are 9,500 vulns. Or none.
'Or none' is ruled out since it found the same vulnerability - I agree that there is a question on precision on the smaller model, but barring further analysis it just feels like '9500' is pure vibes from yourself? Also (out of interest) did Anthropic post their false-positive rate?
The smaller model is clearly the more automatable one IMO if it has comparable precision, since it's just so much cheaper - you could even run it multiple times for consensus.
Admittedly just vibes from me, having pointed small models at code and asked them questions, no extensive evaluation process or anything. For instance, I recall models thinking that every single use of `eval` in javascript is a security vulnerability, even something obviously benign like `eval("1 + 1")`. But then I'm only posting comments on HN, I'm not the one writing an authoritative thinkpiece saying Mythos actually isn't a big deal :-)
> 'Or none' is ruled out since it found the same vulnerability
It's not, though. It wasn't asked to find vulnerabilities over 10,000 files - it was asked to find a vulnerability in the one particular place in which the researchers knew there was a vulnerability. That's not proof that it would have found the vulnerability if it had been given a much larger surface area to search.
Yeah. And it is totally depressing that this article got voted to the top of the front page. It means people aren’t capable of this most basic reasoning so they jumped on the “aha! so the mythos announcement was just marketing!!”
> because small models found the same vulnerability.
With a ton of extra support. Note this key passage:
>We isolated the vulnerable svc_rpc_gss_validate function, provided architectural context (that it handles network-parsed RPC credentials, that oa_length comes from the packet), and asked eight models to assess it for security vulnerabilities.
Yeah it can find a needle in a haystack without false positives, if you first find the needle yourself, tell it exactly where to look, explain all of the context around it, remove most of the hay and then ask it if there is a needle there.
It's good for them to continue showing ways that small models can play in this space, but in my read their post is fairly disingenuous in saying they are comparable to what Mythos did.
I mean this is the start of their prompt, followed by only 27 lines of the actual function:
> You are reviewing the following function from FreeBSD's kernel RPC subsystem (sys/rpc/rpcsec_gss/svc_rpcsec_gss.c). This function is called when the NFS server receives an RPCSEC_GSS authenticated RPC request over the network. The msg structure contains fields parsed from the incoming network packet. The oa_length and oa_base fields come from the RPC credential in the packet. MAX_AUTH_BYTES is defined as 400 elsewhere in the RPC layer.
The original function is 60 lines long, they ripped out half of the function in that prompt, including additional variables presumably so that the small model wouldn't get confused / distracted by them.
You can't really do anything more to force the issue except maybe include in the prompt the type of vuln to look for!
It's great they they are trying to push small models, but this write up really is just borderline fake. Maybe it would actually succeed, but we won't know from that. Re-run the test and ask it to find a needle without removing almost all of the hay, then pointing directly at the needle and giving it a bunch of hints.
What the source article claims is that small models are not uniformly worse at this, and in fact they might be better at certain classes of false positive exclusion. This is what Test 1 seems to show.
(I would emphasize that the article doesn't claim and I don't believe that this proves Mythos is "fake" or doesn't matter.)
Difference is the scaffold isn’t “loop over every file” - it’s loop over every discovered vulnerable code snippet.
If you isolate the codebase just the specific known vulnerable code up front it isn’t surprising the vulnerabilities are easy to discover. Same is true for humans.
Better models can also autonomously do the work of writing proof of concepts and testing, to autonomously reject false positives.
> I hypothesize it will find the exploit, but it will also turn up so much irrelevant nonsense that it won't matter.
The trick with Mythos wasn't that it didn't hallucinate nonsense vulnerabilities, it absolutely did. It was able to verify some were real though by testing them.
The question is if smaller models can verify and test the vulnerabilities too, and can it be done cheaper than these Mythos experiments.
People often undervalue scaffolding. I was looking at a bug yesterday, reported by a tester. He has access to Opus, but he's looking through a single repo, and Amazon Q. It provided some useful information, but the scaffolding wasn't good enough.
I took its preliminary findings into Claude Code with the same model. But in mine it knows where every adjacent system is, the entire git history, deployment history, and state of the feature flags. So instead of pointing at a vague problem, it knew which flag had been flipped in a different service, see how it changed behavior, and how, if the flag was flipped in prod, it'd make the service under testing cry, and which code change to make to make sure it works both ways.
It's not as if a modern Opus is a small model: Just a stronger scaffold, along with more CLI tools available in the context.
The issue here in the security testing is to know exactly what was visible, and how much it failed, because it makes a huge difference. A middling chess player can find amazing combinations at a good speed when playing puzzle rush: You are handed a position where you know a decisive combination exist, and that it works. The same combination, however, might be really hard to find over the board, because in a typical chess game, it's rare for those combinations to exist, and the energy needed to thoroughly check for them, and calculate all the way through every possible thing. This is why chess grandmasters would consider just being able to see the computer score for a position to be massive cheating: Just knowing when the last move was a blunder would be a decisive advantage.
When we ask a cheap model to look for a vulnerability with the right context to actually find it, we are already priming it, vs asking to find one when there's nothing.
Calling it “expert orchestration” is misleading when they were pointing it at the vulnerable functions and giving it hints about what to look for because they already knew the vulnerability.
You know for loops exist and you can run opencode against any section of code with just a small amount of templating, right? There's zero stopping you from writing a harness that does what you're saying.
OTOH, this article goes too far the opposite extreme:
> We isolated the vulnerable svc_rpc_gss_validate function, provided architectural context (that it handles network-parsed RPC credentials, that oa_length comes from the packet), and asked eight models to assess it for security vulnerabilities.
To follow your analogy, they pointed to the exact room where the gold was hidden, and their model found it. But finding the right room within the entire continent in honestly the hard part.
That was my thought exactly. If small models can find these same vulnerabilities, and your company is trying to find vulnerabilities, why didn’t you find them?
I speculatively fired Claude Opus 4.6 at some code I knew very well yesterday as I was pondering the question. This code has been professionally reviewed about a year ago and came up fairly clean, with just a minor issue in it.
Opus "found" 8 issues. Two of them looked like they were probably realistic but not really that big a deal in the context it operates in. It labelled one of them as minor, but the other as major, and I'm pretty sure it's wrong about it being "major" even if is correct. Four of them I'm quite confident were just wrong. 2 of them would require substantial further investigation to verify whether or not they were right or wrong. I think they're wrong, but I admit I couldn't prove it on the spot.
It tried to provide exploit code for some of them, none of the exploits would have worked without some substantial additional work, even if what they were exploits for was correct.
In practice, this isn't a huge change from the status quo. There's all kinds of ways to get lots of "things that may be vulnerabilities". The assessment is a bigger bottleneck than the suspicions. AI providing "things that may be an issue" is not useless by any means but it doesn't necessarily create a phase change in the situation.
An AI that could automatically do all that, write the exploits, and then successfully test the exploits, refine them, and turn the whole process into basically "push button, get exploit" is a total phase change in the industry. If it in fact can do that. However based on the current state-of-the-art in the AI world I don't find it very hard to believe.
It is a frequent talking point that "security by obscurity" isn't really security, but in reality, yeah, it really is. An unknown but presumably staggering number of security bugs of every shape and size are out there in the world, protected solely by the fact that no human attacker has time to look at the code. And this has worked up until this point, because the attackers have been bottlenecked on their own attention time. It's kind of just been "something everyone knows" that any nation-state level actor could get into pretty much anything they wanted if they just tried hard enough, but "nation-state level" actor attention, despite how much is spent on it, has been quite limited relative to the torrent of software coming out in the world.
Unblocking the attackers by letting them simply purchase "nation-state level actor"-levels of attention in bulk is huge. For what such money gets them, it's cheap already today and if tokens were to, say, get an order of magnitude cheaper, it would be effectively negligible for a lot of organizations.
In the long run this will probably lead to much more secure software. The transition period from this world to that is going to be total chaos.
... again, assuming their assessment of its capabilities is accurate. I haven't used it. I can't attest to that. But if it's even half as good as what they say, yes, it's a huge huge huge deal and anyone who is even remotely worried about security needs to pay attention.
Maybe they did use small models but you couldn't make the front page of HN with something like this until Anthropic made a big fuss out of it. Or perhaps it is just a question of compute. Not everyone has 20k$ or the GPU arsenal to task models to find vulnerabilities which may/may not be correct?
Unless Anthropic makes it known exactly what model + harness/scaffolding + prompt + other engineering they did, these comparisons are pointless. Given the AI labs' general rate of doomsday predictions, who really knows?
papers are always coming out saying smaller models can do these amazing and terrifying things if you give them highly constrained problems and tailored instructions to bias them toward a known solution. most of these don't make the front page because people are rightfully unimpressed
It seems feasible to use a small/cheap model to flag possible vulnerabilities, and then use a more expensive model to do a second-pass to confirm those, rather than on every file. Could dramatically reduce the total cost and speed up the process.
This is a really interesting point though -- it's really scaffold-dependent.
Because for the same price, you could point the small model at each function, one by one, N times each, across N prompts instructing it to look for a specific class of issue.
It's not that there's no difference between models, but it's hard to judge exactly how much difference there is when so much depends on the scaffold used. For a properly scientific test, you'd need to use exactly the same one.
Which isn't possible when Anthropic won't release the model.
How is that a direct comparison? The link you gave has a quote that says it’s not:
> Scoped context: Our tests gave models the vulnerable function directly, often with contextual hints (e.g., "consider wraparound behavior"). A real autonomous discovery pipeline starts from a full codebase with no hints
They pointed the models at the known vulnerable functions and gave them a hint. The hint part is what really breaks this comparison because they were basically giving the model the answer.
How much of that is simply scale? Anthropic threw probably an entire data center at analyzing a code base. Has anyone done the same with a "small" model?
I'm having trouble finding this info (I assume they won't publish it), but could the secret sauce be much larger and more readily accessible context window?
OpenBSD's code is in the 10s of millions of lines. Being able to hold all of it in context would make bug finding much easier.
You can look at some of the bugs, if you'd like. They are (at least the ones I looked at) fairly self-contained, scoped to a single function, a hundred lines or less. There's no need for a massive amount of context.
We don't even need to hypothesize that much on the irrelevant nonsense, since they helpfully provide data with the detected vulnerability patched: https://aisle.com/blog/ai-cybersecurity-after-mythos-the-jag... and half of the small models they touted as finding the vulnerability still found it in the patched code in 3/3 runs. A model that finds a vulnerability 100% of the time even when there is none is just as informative as a model that finds a vulnerability 0% of the time even when there is one. You could replace it with a rock that has "There's a vulnerability somewhere." engraved on it.
They're a company selling a system for detecting vulnerabilities reliant on models trained by others, so they're strongly incentivized to claim that the moat is in the system, not the model, and this post really puts the thumb on the scale. They set up a test that can hardly distinguish between models (just three runs, really??) unless some are completely broken or work perfectly, the test indeed suggests that some are completely broken, and then they try to spin it as a win anyway!
A high false-positive rate isn't necessarily an issue if you can produce a working PoC to demonstrate the true positives, where they kinda-sorta admit that you might need a stronger model for this (a.k.a. what they can't provide to their customers).
Overall I rate Aisle intellectually dishonest hypemongers talking their own book.
Can't you execute the bug to see if the vulnerability is real? So you have a perfect filter. Maybe Mythos decided w/o executing but we don't know that.
What he's saying is that you should read the "Caveats and limitations" section of the article.
Here's the first one:
> Our tests gave models the vulnerable function directly, often with contextual hints (e.g., "consider wraparound behavior").
Mythos did no such thing, it was cut lose and told to find vulnerabilities. If the intent was to prove that small models are just as good, they haven't demonstrated that at all. The end.
> We took the specific vulnerabilities Anthropic showcases in their announcement, isolated the relevant code, and ran them through small, cheap, open-weights models. Those models recovered much of the same analysis. Eight out of eight models detected Mythos's flagship FreeBSD exploit, including one with only 3.6 billion active parameters costing $0.11 per million tokens.
Impressive, and very valuable work, but isolating the relevant code changes the situation so much that I'm not sure it's much of the same use case.
Being able to dump an entire code base and have the model scan it is they type of situation where it opens up vulnerability scans to an entirely larger class of people.
This is from the first of the caveats that they list:
> Scoped context: Our tests gave models the vulnerable function directly, often with contextual hints (e.g., "consider wraparound behavior"). A real autonomous discovery pipeline starts from a full codebase with no hints. The models' performance here is an upper bound on what they'd achieve in a fully autonomous scan. That said, a well-designed scaffold naturally produces this kind of scoped context through its targeting and iterative prompting stages, which is exactly what both AISLE's and Anthropic's systems do.
That's why their point is what the subheadline says, that the moat is the system, not the model.
Everybody so far here seems to be misunderstanding the point they are making.
If that's the point they are making, let's see their false positive rate that it produces on the entire codebase.
They measured false negatives on a handful of cases, but that is not enough to hint at the system you suggest. And based on my experiences with $$$ focused eval products that you can buy right now, e.g. greptile, the false positive rate will be so high that it won't be useful to do full codebase scans this way.
I get what you're saying, but I think this is still missing something pretty critical.
The smaller models can recognize the bug when they're looking right at it, that seems to be verified. And with AISLE's approach you can iteratively feed the models one segment at a time cheaply. But if a bug spans multiple segments, the small model doesn't have the breadth of context to understand those segments in composite.
The advantage of the larger model is that it can retain more context and potentially find bugs that require more code context than one segment at a time.
That said, the bugs showcased in the mythos paper all seemed to be shallow bugs that start and end in a single input segment, which is why AISLE was able to find them. But having more context in the window theoretically puts less shallow bugs within range for the model.
I think the point they are making, that the model doesn't matter as much as the harness, stands for shallow bugs but not for vulnerability discovery in general.
OK, consider a for loop that goes through your repo, then goes through each file, and then goes through each common vulnerability...
Is Mythos some how more powerful than just a recursive foreloop aka, "agentic" review. You can run `open code run --command` with a tailored command for whatever vulnerabilities you're looking for.
newer models have larger context windows, and more stable reasoning across larger context windows.
If you point your model directly at the thing you want it to assess, and it doesn't have to gather any additional context you're not really testing those things at all.
Say you point kimi and opus at some code and give them an agentic looping harness with code review tools. They're going to start digging into the code gathering context by mapping out references and following leads.
If the bug is really shallow, the model is going to get everything it needs to find it right away, neither of them will have any advantage.
If the bug is deeper, requires a lot more code context, Opus is going to be able to hold onto a lot more information, and it's going to be a lot better at reasoning across all that information. That's a test that would actually compare the models directly.
Mythos is just a bigger model with a larger context window and, presumably, better prioritization and stronger attention mechanisms.
Harnesses are basically doing this better than just adding more context. Every time, REGARDLESS OF MODEL SIZE, you add context, you are increasing the odds the model will get confused about any set of thoughts. So context size is no longer some magic you just sprinkle on these things and they suddenly dont imagine things.
So, it's the old ML join: It's just a bunch of if statements. As others are pointing out, it's quite probably that the model isn't the thing doing the heavy lifting, it's the harness feeding the context. Which this link shows that small models are just as capabable.
Which means: Given a appropiately informed senior programmer and a day or two, I posit this is nothing more spectacular than a for loop invoking a smaller, free, local, LLM to find the same issues. It doesn't matter what you think about the complexity, because the "agentic" format can create a DAG that will be followable by a small model. All that context you're taking in makes oneshot inspections more probable, but much like how CPUs have go from 0-5 ghz, then stalled, so too has the context value.
Agent loops are going to do much the same with small models, mostly from the context poisoning that happens every time you add a token it raises the chance of false positives.
To clarify, I don't necessarily agree with the post or their approach. I just thought folks were misreading it. I also think it adds something useful to the conversation.
> That's why their point is what the subheadline says, that the moat is the system, not the model.
I'm skeptical; they provided a tiny piece of code and a hint to the possible problem, and their system found the bug using a small model.
That is hardly useful, is it? In order to get the same result , they had to know both where the bug is and what the bug is.
All these companies in the business of "reselling tokens, but with a markup" aren't going to last long. The only strategy is "get bought out and cash out before the bubble pops".
You can imagine a pipeline that looks at individual source files or functions. And first "extracts" what is going on. You ask the model:
- "Is the code doing arithmetic in this file/function?"
- "Is the code allocating and freeing memory in this file/function?"
- "Is the code the code doing X/Y/Z? etc etc"
For each question, you design the follow-up vulnerability searchers.
For a function you see doing arithmetic, you ask:
- "Does this code look like integer overflow could take place?",
For memory:
- "Do all the pointers end up being freed?"
_or_
- "Do all pointers only get freed once?"
I think that's the harness part in terms of generating the "bug reports". From there on, you'll need a bunch of tools for the model to interact with the code. I'd imagine you'll want to build a harness/template for the file/code/function to be loaded into, and executed under ASAN.
If you have an agent that thinks it found a bug: "Yes file xyz looks like it could have integer overflow in function abc at line 123, because...", you force another agent to load it in the harness under ASAN and call it. If ASAN reports a bug, great, you can move the bug to the next stage, some sort of taint analysis or reach-ability analysis.
So at this point you're running a pipeline to:
1) Extract "what this code does" at the file, function or even line level.
2) Put code you suspect of being vulnerable in a harness to verify agent output.
3) Put code you confirmed is vulnerable into a queue to perform taint analysis on, to see if it can be reached by attackers.
Traditionally, I guess a fuzzer approached this from 3 -> 2, and there was no "stage 1". Because LLMs "understand" code, you can invert this system, and work if up from "understanding", i.e. approach it from the other side. You ask, given this code, is there a bug, and if so can we reach it?, instead of asking: given this public interface and a bunch of data we can stuff in it, does something happen we consider exploitable?
That's funny, this is how I've been doing security testing in my code for a while now, minus the 'taint analysis'. Who knew I was ahead of the game. :P
In all seriousness though, it scares me that a lot of security-focused people seemingly haven't learned how LLMs work best for this stuff already.
You should always be breaking your code down into testable chunks, with sets of directions about how to chunk them and what to do with those chunks. Anyone just vaguely gesturing at their entire repo going, "find the security vulns" is not a serious dev/tester; we wouldn't accept that approach in manual secure coding processes/ SSDLCs.
Tunnel vision? If your model can handle big context, why divide into lesser problems to conquer - even if such splitting might be quite trivial and obvious?
It's the difference of "achieve the goal", and "achieve the goal in this one particular way" (leverage large context).
I meant, if the claim here is that small models can accomplish the same things with good scaffolding, why didn’t they demonstrate finding those problem with good scaffolding rather than directly pointing them at the problem?
> Anthropic's own scaffold is described in their technical post: launch a container, prompt the model to scan files, let it hypothesize and test, use ASan as a crash oracle, rank files by attack surface, run validation. That is very close to the kind of system we and others in the field have built, and we've demonstrated it with multiple model families, achieving our best results with models that are not Anthropic's. The value lies in the targeting, the iterative deepening, the validation, the triage, the maintainer trust. The public evidence so far does not suggest that these workflows must be coupled to one specific frontier model.
The argument in the article is that the framework to run and analyze the software being tested is doing most of the work in Anthropic's experiment, and that you can get similar results from other models when used in the same way.
The thing is with smaller cheaper models it is very possible to simply take every file in a codebase, and prompt it asking for it to find vulnerabilities.
You could even isolate it down to every function and create a harness that provides it a chain of where and how the function is used and repeat this for every single function in a codebase.
For some very large codebases this would be unreasonable, but many of the companies making these larger models do realistically have the compute available to run a model on every single function in most codebases.
You have the harness run this many times per file/function, and then find ones that are consistently/on average pointed as as possible vulnerability vectors, and then pass those on to a larger model to inspect deeper and repeat.
Most of the work here wouldn't be the model, it'd be the harness which is part of what the article alludes to.
> it is very possible to simply take every file in a codebase, and prompt it asking for it to find vulnerabilities.
My understanding (based on the Security, Cryptography, Whatever podcast interview[0] -- which, by the way, go listen to it) is that this is actually what Anthropic did with the large model for these findings.
> I wrote a single prompt, which was the same for all of the content management systems, which is, I would like you to audit the security of this codebase. This is a CMS. You have complete access to this Docker container. It is running. Please find a bug. And then I might give a hint. “Please look at this file.” And I’ll give different files each time I invoke it in order to inject some randomness, right? Because the model is gonna do roughly the same time each time you run it. And so if I want to have it be really thorough, instead of just running 100 times on the same project, I’ll run it 100 times, but each time say, “Oh, look at this login file, look at this other thing.” And just enumerate every file in the project basically.
Isn't the difference just harness then? I can write a harness that chunks code into individual functions or groups of functions and then feed it into a vulnerability analysis agent.
It's probably not the 'only' difference, because clearly the models are advancing in capability, but it's likely way more important than generally given credit for.
If you cut out the vulnerable code from Heartbleed and just put it in front of a C programmer, they will immediately flag it. It's obvious. But it took Neel Mehta to discover it. What's difficult about finding vulnerabilities isn't properly identifying whether code is mishandling buffers or holding references after freeing something; it's spotting that in the context of a large, complex program, and working out how attacker-controlled data hits that code.
No, writing an advertisement is not weird. What's weird is that it's top of HN. Or really, no, this isn't weird either if you think about it -- people lookin for a gotcha "Oh see, that new model really isn't that good/it's surely hitting a wall/plateau any day now" upvoted it.
It's also that humans are very bad at repetitive detailed tasks. Sitting down with a code base and looking at each function for integer overflow comparison bugs gets boring really fast. It's a rare person who can do that for as long as it takes to find a bug that they don't already have some clues about.
It's the flaw in the "given enough eyeballs, all bugs are shallow" argument. Because eyeballs grow tired of looking at endless lines of code.
Machines on the other hand are excellent at this. They don't get bored, they just keep doing what they are told to do with no drop-off in attention or focus.
idk man, pay me enough money and I’ll look at as much code as you want looking for integer overflows
Would it be cheaper than Claude Mythos doing it? No idea. Maybe, maybe not.
But it’s weird how we’re willing to throw away money to a megacorp to do it with “automation” for potentially just as much if not more as it would cost to just have big bounty program or hiring someone for nearly the same cost and doing it “normally”.
It would really have to be substantially less cost for me to even consider doing it with a bot.
> idk man, pay me enough money and I’ll look at as much code as you want looking for integer overflows
So would I, but it doesn't negate that we, humans, are bad at this. We will get bored and our focus will begin to drift. We might not notice it, we might not want to admit it, but after a few continuous hours we will start missing things.
It's weird, because when working on a big project, taking a break for a week or two, and returning to it, I will find a bug and will see hundreds of lines of code that are absolutely terrible, and I will tell myself "Tom you know better than to do this, this is a rookie mistake".
I think people forget that it's hard to be clever and tidy 100% of the time. Big programs take a lot of discipline and an understanding of the context that can be really hard to maintain. This is one of several reasons that my second draft or third draft of code is almost always considerably better than the first draft.
If it’s obvious when you look close, then automate looking close. Seems simple to write tools that spider thru a code base, finding logical groupings and feeding them into an LLM with prompts like “there is a vulnerability in this code, find it”.
The thesis is, the tooling is what matters - the tools (what they call the harness) can turn a dumb llm into a smart llm.
Hold on, I misread your comment because I'm knee-jerk about code scanners, which were the bane of my existence for a while. Reworking... and: done. The original comment was just the first graf without the LLM qualification. Sorry about that.
The general approach without LLMs doesn't work. 50 companies have built products to do exactly what you propose here; they're called static application security testing (SAST) tools, or, colloquially, code scanners. In practice, getting every "suspicious" code pattern in a repository pointed out isn't highly valuable, because every codebase is awash in them, and few of them pan out as actual vulnerabilities (because attacker-controlled data never hits them, or because the missing security constraint is enforced somewhere else in the call chain).
Could it work with LLMs? Maybe? But there's a big open question right now about whether hyperspecific prompts make agents more effective at finding vulnerabilities (by sparing context and priming with likely problems) or less effective (by introducing path dependent attractors and also eliminating the likelihood of spotting vulnerabilities not directly in the SAST pattern book).
I have long said that static checkers get ten false positives. note that size of the code is not a consideration, it doesn't matter if it the four line 'hello world' or the 10 million line monster some of us work on, it is ten max false positive.
The point of contention is whether Mythos is the product of its intelligence or its harness; the results like this, and other similar testimonies, call into question too-dangerous-to-release marketing, and for good reason, too. Because it is powerful marketing. Aisle merely says the intelligence is there in the small models. I say, it's already clear that competent defenders could viably mimic, or perhaps even eclipse what Mythos does, by (a) making better harness, (b) simply spending more on batch jobs, bootstrapping, cache better, etc. You may not be doing this yourself, but your probably should.
The technique Anthropic uses was demonstrated by Nicholas Carlini in a talk he gave 2 weeks ago and it's very simple, when asking LLMs to review code, ask them to focus its review on one file in a single session. Here is the video with the timestamp (watch through to ~5:30, they show two different ways of prompting claude).
IMO the big "innovation" being shown by Mythos is the effectiveness with prompting LLMs to look for security vulnerabilities by focusing on specific files one at a time and automating this prompting with a simple script.
Prompting Mythos to focus on a single file per session is why I suspect it cost Anthropic $20k to find some of the bugs in these codebases. I know this same technique is effective with Opus 4.6 and GPT 5.4 because I've been using it on my own code. If you just ask the agent to review your pr with a low effort prompt they are not exhaustive, they will not actually read each changed file and look at how it interacts with the system as a whole. If the entire session is to review the changes for a single file, the llm will do much more work reviewing it.
Edit: I changed my phrasing, it's not about restricting its entire context to one file but focusing it on one file but still allowing it to look at how other files interact with it.
Congrats: completely broken methodology, with a big conflict of interest. Giving specific bug hints, with an isolated function that is suspected to have bugs, is not the same task, NOR (crucially) is a task you can decompose the bigger task into. It is basically impossible to segment code in pieces, provide pieces to smaller models, and expect them to find all the bugs GPT 5.4 or other large models can find. Second: the smarter the model, and less the pipeline is important. In the latest couple of days I found tons if Redis bugs with a three prompts open-ended pipeline composed of a couple of shell scripts. Do you think I was not already tying with weaker models? I did, but it didn't work. Don't trust what you read, you have access to frontier models for 20$ a month. Download some C code, create a trivial pipeline that starts from a random file and looks for vulnerabilities, then another step that validates it under a hard test, like ASAN crash, or ability to reach some secret, and so forth, and only then the problem can be reported. Test yourself what it is possible. Don't let your fear make you blind. Also, there is a big problem that makes the blog post reasoning not just weak per se, but categorically weak: if small model X can find 80% of vulnerabilities, if there is a model Y that can find the other potential 20%, we need "Y": the maintainers should make sure they access to models that are at least as good as the black hats folks.
Exactly, this is so flawed. Anthropic themselves said they only reported <1% of the vulnerabilities found, cause the rest is unpatched.
Give open models an environment (prior to Feb 15- so no Mythos-discovered vulns are patche) of Linux and see how many vulnerabilities it can find. Then put it in a sandbox and see if it can escape and send you an e-mail.
> Those models recovered much of the same analysis
This is an essentially unquantifiable statement that makes the underlying claim harder to believe as an external party. What does “much” mean here? The end state of vulnerability exploitation is typically eminently quantifiable (in the form of a functional PoC that demonstrates an exploited end state), so the strong version of the claims here would ideally be backed up by those kinds of PoCs.
(Like other readers, I also find the trick of pre-feeding the smaller models the “relevant” code to be potentially disqualifying in a fair comparison. Discovering the relevant code is arguably one of the hardest parts of human VR.)
Without showing false-positive rates this analysis is useless.
If your model says every line if your code has a bug, it will catch 100% of the bugs, but it's not useful at all. They tested false-positives with only a single bug...
I'm not defending anthropic and openai either. Their numbers are garbage too since they don't produce false-positive rates either.
Yes, and in this case they pointed at the function, so a 1-bit model ("yes") would be correct. But it's not that bad. First, they included a test with a false positive. The small models got it right, Opus got it wrong. Second, they asked for an analysis. Look for "Exploitation reasoning, single follow-up prompt:" in the post. It's hard to tell how good they were at a glance, though apparently the full logs are available so you could pull them up.
Anyway, it seems like they erred in the up-front claim "small models found the vulnerability we pointed directly at!", but the findings are at least somewhat stronger if you read through the details.
The small models didn't match Mythos at exploitation. They suggested plausible exploits, but didn't actually try them out so I can't tell if they would have worked. Deepseek R1's sounds pretty convincing to me, but I'm not a good judge. (I'm more in the space of accidentally writing vulnerabilities, not seeking them out or exploiting them. Well, ok, I have a static analysis that finds some, at least.)
Their isolation approach is totally different from Mythos approach though. Mythos had to evaluate whole code bases rather than isolated sections. It's like saying one dog walked into the Amazon jungle and found a tennis ball and then another team isolated a 1 square kilometer radius that they knew the ball was definitely in and found the same ball.
Even that would be more meaningful test. They basically coated the ball with a strong smell, then they prepped the dog with that smell, then set it loose in a 5x5 meter area.
"Our tests gave models the vulnerable function directly, often with contextual hints (e.g., "consider wraparound behavior")."
I don’t think mythos can ingest an entire codebase into context. So it’s spinning off sub-agents to process chunks. Which supports their thesis: the harness is the moat. The tooling is whats important, the model is far far less important.
Mythos was clear it was one agent per chunk. But this positive confirming results do not actually disprove anytime with Mythos, because it is only one side of the discriminator challenge - you got positives, but we do not know your false positive rate and your false negative rate.
If smaller models can find these things, that doesn’t mean mythos is worse than we thought. It means all models are more capable.
Also if pointing models at files and giving them hints is all it takes to make them find all kinds of stuff, well, we can also spray and pray that pretty well with llms can’t we.
It just points to us finding a lot more stuff with only a little bit more sophistication.
Hopefully the growing pains are short and defense wins
The best way to think of Anthropic's communication about Mythos is as advertisement. It's basically "our model is too smart to release" which suggests they're ahead of OpenAI (without proof)
The whole company is like that. If things were as amazing as advertised, they wouldn't even need to advertise. Or to release models to the public at all.
The impact of the Mythos announcement on the cybersecurity firms( like Crowdstrike,ZScalar etc) is big enough(10-15% drop in stock price) and this pushback is expected.
Companies like Aisle.com (the blog) and other VAPT companies charge huge amounts to detect vulnerabilities.
If Cloud Mythos become a simple github hook their value will get reduced.
LLMs are wordsmith oracles. A lot of effort went into trying to coax interactive intelligence from them but the truth is that you could have probably always harnessed the base models directly to do very useful things. The instruct tuned models give your harness even more degrees of freedom.
A while ago, the autoresearch[1] harness went viral, yet it's but a highly simplified version of AlphaEvolve[2][3][4].
In the cybersecury context, you can envision a clever harness that probes every function in a codebase for vulnerabilities, then bubbles the candidates up to their callsites (and probes whether the vulnerability can be triggered from there) and then all the way to an interface (such as a syscall) where a potential exploit can be manifested. And those would be the low hanging fruit, other vulnerabilities may require the interplay of multiple functions. Or race conditions.
1. Mythos uniquely is able to find vulnerabilities that other LLMs cannot practically.
2. All LLMs could already do this but no one tried the way anthropic did.
The truth is one of these. And it comes down whether the comparison is apples to apples. Since we don't know the exact specifics of how either tests were performed, we lack a way of knowing absolutely.
So I guess, like so many things today, we can to pick the truth we find most comfortable personally.
My theory is that Mythos is basically just Opus with revised context window handling and more compute thrown at it. So while it will be a step forward, it is probably primarily hype.
If you isolate the positive cases and then ask a tool to label them and it labels them all positive, doesn't prove anything. This is a one-sided test and it is really easy to write a tool that passes it -- just return always true!
You need to test your tool on both positive and negative cases and check if it is accurate on both.
If you don't, you could end up with hundreds or thousands of false positives when using this on real-world samples.
The real test is to use it to find new real bugs in the midst of a large code base.
This misses the broader ongoing trend. For a few million dollars, of course you can create a startup that builds tools it can use to more efficiently find code vulnerabilities. And of course you can do this with weaker models with scaffolds that incorporate lots of human understanding. The difference now is that you don't need an expensive team, nor a bunch of human heuristics, nor a million dollars. The requisite cost and skill are falling rapidly.
They did do one agent per code chunk, yes. But key is that their agent had to identify when there was a vulnerability and when there wasn't. This "small model" test only had to label the known positive cases as positive -- which any function that simply returns "true" can do. This whole test setup is annoying because it proves nothing.
to be fair, last post i saw from anthropic on finding linux kernel vulnerability was a while loop per failed prompting "there is a vulnerability here, find it"
more important than that, no frontier model can keep the entire linux kernel in context, so there definitely is code isolation, either explicitly or implicitly (the model itself delegates subagents with smaller chunks of code)
Didn’t they also use Mythos to scan Linux many times over and it only found one DoS bug or something? I find it hard to believe there is only one security bug lurking.
There are a lot of details in the original article, in most cases comparing with Opus, which required "human guidance" to exploit the FreeBSD vulnerability:
Intuitively every existing model has already been trained on all code, all vulnerabilities reported, all security papers. So they all have the capability. Small models fall short because they may not be able to find a vulnerability that spans across a large function chain but for the most part they should suffice too.
Of course I say this without any knowledge of what mythos is doing or how it’s different. I am sure it’s somehow different
Not intuitive at all. Not all models are equally capable, just because they had the same training data. The model architecture (as a whole) is very important. To reduce capability, you can reduce layers, tool use, thinking, quantize it, etc. This is trivially proven by a cursory glance in the rough direction of any set of benchmarks (or actual use).
Using small models as a classifier "there might be a vulnerability here" is probably reasonable, if you have a model capable of proving it. There are many companies attempting this without the verification step, resulting in AI vulnerability checker being banned left and right, from the nonsense noise.
Wouldn't this mean we're even more cooked? I've seen this page cited a few times as evidence that Mythos is no big deal, but if true then the same big deal is already out there with other models today.
This would just speed up the discovery -> patch cycle, at least until such time that all the low hanging fruit (=represented in training data) is patched.
Though another possibility would be that since LLMs generate so much code, the LLM vulnerability discovery would just keep chugging along and we'd simply settle for the same amount of potential vulns, same relative vulnerability-exploit-patch dynamics, though higher in absolute numbers.
finding vulns in a large codebase is a search problem with a huge
negative space and what aisle measured is classification accuracy on
ground-truth positives, those are different tasks so a model that
correctly labels a pre-isolated vulnerable function tells me almost
nothing about that model's ability to surface the same function out of
a million lines of unrelated code under a realistic triage budget
the experiment i'd want to see is running each of the small models as an
unsupervised scanner across full freebsd then return the top-k suspicious
functions per model and compute precision at recall levels that
correspond to real analyst triage budgets, if mythos s findings show up
in the small models top 100, i'd call that meaningful but if they only
surface under 10k false positives then the cost advantage
collapses because analyst triage time is more expensive than frontier
model compute to begin with
second thing i keep coming back to is the $20k mythos number is a search
budget not a model cost, small models at one hundredth the per-token
price don't give us one hundredth the total budget when the search
process is the same shape, i still run thousands of iterations and the
issue for autonomous vuln research is how fast the reward signal
converges and the aisle post doesn't touch any of this
The methodology here is completely wrong, outright dishonest.
Finding a needle in a haystack is easy if someone hands you the small handful of hay containing the needle up front, and raises their eyebrows at you saying “there might be a needle in this clump of hay”.
I think that probably Mytho's mojo comes from a lot of post-training on this kind of task.
I occasionally pick up contract work doing coding annotation to make some quick extra money, and a few months ago one of the projects was heavily focused on spotting common memory access bugs in C and C++.
I don't dispute the fact that it's more than cool that we have a new tool to find security exploits (and do many other things) but... A big shoot-out to OpenBSD?
We're literally talking about the biggest computers on the planet ever, trained with the biggest amount of data ever available to a system, with the biggest investment ever made by man or close to it and...
The subtlest security bug it can find required: going 28 years in the past and find a...
Denial-of-service?
A freaking DoS? Not a remote root exploit. Not a local exploit.
Just a DoS? And it had to go into 28 years old code to find that?
So kudos, hats off, deep bow not to Mythos but to OpenBSD? Just a bit, no!?
I mean isn't that most of it? If you put a snippet of code in front of me and said "there's probably a vulnerability here" I could probably spend a few hours (a much lower METR time!) and find it. It's a whole other ballgame to ask me with no context to come up with an exploit.
Sure. But it’s a computer. You can run “there’s probably a vulnerability here” as many times as you like. And it’s easier and cheaper to run it many times with a small open model than a big frontier model.
It also sounds like that is how mythos works too. Which makes sense - the linux kernel is too big to fit in context
At the center of every security situation is the question, "is the effort worth the reward?"
We prepare security measures based on the perceived effort a bad actor would need to defeat that method, along with considering the harm of the measure being defeated. We don't build Fort Knox for candy bars, it was built for gold bars.
These model advances change the equation. The effort and cost to defeat a measure goes down by an order of magnitude or more.
Things nobody would have considered to reasonably attempt are becoming possible. However. We have 2000-2020s security measures in place that will not survive the AI models of 2026+. The investment to resecure things will be massive, and won't come soon enough.
Mythos is clearly a nice improvement. It’s also clear there’s a lot of unfounded hype around it to keep the AI hype cycle going.
Gating access is also a clever marketing move:
Option A: Release it but run out of capacity, everyone is annoyed and moves on. Drives focus back to smaller models.
Option B: A bunch of manufactured hype and putting up velvet ropes around it saying it’s “too dangerous” to let near mortals touch it. Press buys it hook, like, and sinker, sidesteps the capacity issues and keeps the hype train going a bit longer.
Seems quite clear we’re seeing “Option B” play out here.
The thesis that the system is more important than the model is not bitter lesson pilled. I would not bet on this in the long term. We will get to the point where you can just tell the model to go find and classify the severity of all security problems with a codebase.
It's strange to me they didn't reduce to PoC so the quantitative part is an apples-to-apples comparison. You don't need any fancy tooling, if you want to do this at home you can do something like below in whatever command line agent and model you like. A while back I did take one bug all the way through remediation just out of curiosity.
"""
Your task is to study the following directive, research coding agent prompting, research the directive's domain best practices, and finally draft a prompt in markdown format to be run in a loop until the directive is complete.
Concept: Iterative review -- study an issue, enumerate the findings, fix each of the findings, and then repeat, until review finds no issues.
<directive>
Your job is to run a security bug factory that produces remediation packages as described below. Design and apply a methodology based on best practices in exploit development, lean manufacturing, threat modeling, and the scientific method. Use checklists, templates, and your own scripts to improve token efficiency and speed. Use existing tools where possible. Use existing research and bug findings for the target and similar codebases to guide your search. Study the target's development process to understand what kind of harness and tools you need for this work, and what will work in this development environment. A complete remediation package includes a readme documenting the problem and recommendations, runnable PoC with any necessary data files, and proposed patch.
Track your work in TODO.md (tasks identified as necessary) LOG.md (chronological list of tasks complete and lessons) and STATUS.md (concise summary of the current work being done). Never let these get more than a few minutes out of date. At each step ensure the repo file tree would make sense to the next engineer, and if not reorganize it. Apply iterative review before considering a task complete.
Your task is to run until the first complete remediation package is ready for user review.
Your target is <repo url>.
The prompt will be run as follows, design accordingly. Once the process starts, it is imperative not to interrupt the user until completion or until further progress is not possible. Keep output at each step to a concise summary suitable for a chat message.
```
while output=$(claude -p "$(cat prompt.md)"); do echo "$output"; echo "$output" | grep -q "XDONEDONEX" && break; done
```
</directive>
Draft the prompt into prompt.md, and apply iterative review with additional research steps to ensure will execute the directive as faithfully as possible.
The whole "this tool is too dangerous to be public" idea reeks of marketing. Just like all the "AI is an existential threat" talk a year ago. These companies are using ideas usually reserved for something like nuclear weapons to make their products look more impressive.
Anthropic claim is not necessarily that Mythos found vulnerabilities that other models couldn't but that it could easily exploit them while previous models failed to do that:
> “Opus 4.6 is currently far better at identifying and fixing vulnerabilities than at exploiting them.” Our internal evaluations showed that Opus 4.6 generally had a near-0% success rate at autonomous exploit development. But Mythos Preview is in a different league. For example, Opus 4.6 turned the vulnerabilities it had found in Mozilla’s Firefox 147 JavaScript engine—all patched in Firefox 148—into JavaScript shell exploits only two times out of several hundred attempts. We re-ran this experiment as a benchmark for Mythos Preview, which developed working exploits 181 times, and achieved register control on 29 more.
If that was normal Opus, then it sounds to me like Mythos could be a big model, instruction tuned, but without all the safety/refusal part of training.
The bigger point of focus is that the enterprise value accrues to assets associated with software production.
What happened to all that nonsense about LLM’s solving physics, science etc? Lmao that certainly is not happening.
The natural home of LLM’s is in relation to software production.
The question is can Anthropic and OAI survive? If OAI can’t make their entry into the ad business work then they will fight over the same territory. Meaning both of their chances of survival drop as Google who is a monster in relation to software production will not only seek to kill them but buy their GPU’s at a discounted price.
> We isolated the vulnerable vc_rpc_gss_validate function, provided architectural context (that it handles network-parsed RPC credentials, that oa_length comes from the packet), and asked eight models to assess it for security vulnerabilities.
The Anthropic writeup addresses this explicitly:
> This was the most critical vulnerability we discovered in OpenBSD with Mythos Preview after a thousand runs through our scaffold. Across a thousand runs through our scaffold, the total cost was under $20,000 and found several dozen more findings. While the specific run that found the bug above cost under $50, that number only makes sense with full hindsight. Like any search process, we can't know in advance which run will succeed.
Mythos scoured the entire continent for gold and found some. For these small models, the authors pointed at a particular acre of land and said "any gold there? eh? eh?" while waggling their eyebrows suggestively.
For a true apples-to-apples comparison, let's see it sweep the entire FreeBSD codebase. I hypothesize it will find the exploit, but it will also turn up so much irrelevant nonsense that it won't matter.
Wasn't the scaffolding for the Mythos run basically a line of bash that loops through every file of the codebase and prompts the model to find vulnerabilities in it? That sounds pretty close to "any gold there?" to me, only automated.
Have Anthropic actually said anything about the amount of false positives Mythos turned up?
FWIW, I saw some talk on Xitter (so grain of salt) about people replicating their result with other (public) SotA models, but each turned up only a subset of the ones Mythos found. I'd say that sounds plausible from the perspective of Mythos being an incremental (though an unusually large increment perhaps) improvement over previous models, but one that also brings with it a correspondingly significant increase in complexity.
So the angle they choose to use for presenting it and the subsequent buzz is at least part hype -- saying "it's too powerful to release publicly" sounds a lot cooler than "it costs $20000 to run over your codebase, so we're going to offer this directly to enterprise customers (and a few token open source projects for marketing)". Keep in mind that the examples in Nicholas Carlini's presentation were using Opus, so security is clearly something they've been working on for a while (as they should, because it's a huge risk). They didn't just suddenly find themselves having accidentally created a super hacker.
> Wasn't the scaffolding for the Mythos run basically a line of bash that loops through every file of the codebase and prompts the model to find vulnerabilities in it? That sounds pretty close to "any gold there?" to me, only automated.
But the entire value is that it can be automated. If you try to automate a small model to look for vulnerabilities over 10,000 files, it's going to say there are 9,500 vulns. Or none. Both are worthless without human intervention.
I definitely breathed a sigh of relief when I read it was $20,000 to find these vulnerabilities with Mythos. But I also don't think it's hype. $20,000 is, optimistically, a tenth the price of a security researcher, and that shift does change the calculus of how we should think about security vulnerabilities.
> But the entire value is that it can be automated. If you try to automate a small model to look for vulnerabilities over 10,000 files, it's going to say there are 9,500 vulns. Or none.
'Or none' is ruled out since it found the same vulnerability - I agree that there is a question on precision on the smaller model, but barring further analysis it just feels like '9500' is pure vibes from yourself? Also (out of interest) did Anthropic post their false-positive rate?
The smaller model is clearly the more automatable one IMO if it has comparable precision, since it's just so much cheaper - you could even run it multiple times for consensus.
Admittedly just vibes from me, having pointed small models at code and asked them questions, no extensive evaluation process or anything. For instance, I recall models thinking that every single use of `eval` in javascript is a security vulnerability, even something obviously benign like `eval("1 + 1")`. But then I'm only posting comments on HN, I'm not the one writing an authoritative thinkpiece saying Mythos actually isn't a big deal :-)
[delayed]
> 'Or none' is ruled out since it found the same vulnerability
It's not, though. It wasn't asked to find vulnerabilities over 10,000 files - it was asked to find a vulnerability in the one particular place in which the researchers knew there was a vulnerability. That's not proof that it would have found the vulnerability if it had been given a much larger surface area to search.
Also, what is $20,000 today can be $2000 next year. Or $20...
See e.g. https://epoch.ai/data-insights/llm-inference-price-trends/
Or $200,000 for consumers when they have to make a profit
>Or none
We already know this is not true, because small models found the same vulnerability.
No, they didn't. They distinguished it, when presented with it. Wildly different problem.
Yeah. And it is totally depressing that this article got voted to the top of the front page. It means people aren’t capable of this most basic reasoning so they jumped on the “aha! so the mythos announcement was just marketing!!”
> because small models found the same vulnerability.
With a ton of extra support. Note this key passage:
>We isolated the vulnerable svc_rpc_gss_validate function, provided architectural context (that it handles network-parsed RPC credentials, that oa_length comes from the packet), and asked eight models to assess it for security vulnerabilities.
Yeah it can find a needle in a haystack without false positives, if you first find the needle yourself, tell it exactly where to look, explain all of the context around it, remove most of the hay and then ask it if there is a needle there.
It's good for them to continue showing ways that small models can play in this space, but in my read their post is fairly disingenuous in saying they are comparable to what Mythos did.
I mean this is the start of their prompt, followed by only 27 lines of the actual function:
> You are reviewing the following function from FreeBSD's kernel RPC subsystem (sys/rpc/rpcsec_gss/svc_rpcsec_gss.c). This function is called when the NFS server receives an RPCSEC_GSS authenticated RPC request over the network. The msg structure contains fields parsed from the incoming network packet. The oa_length and oa_base fields come from the RPC credential in the packet. MAX_AUTH_BYTES is defined as 400 elsewhere in the RPC layer.
The original function is 60 lines long, they ripped out half of the function in that prompt, including additional variables presumably so that the small model wouldn't get confused / distracted by them.
You can't really do anything more to force the issue except maybe include in the prompt the type of vuln to look for!
It's great they they are trying to push small models, but this write up really is just borderline fake. Maybe it would actually succeed, but we won't know from that. Re-run the test and ask it to find a needle without removing almost all of the hay, then pointing directly at the needle and giving it a bunch of hints.
The prompt they used: https://github.com/stanislavfort/mythos-jagged-frontier/blob...
Compare it to the actual function that's twice as long.
What the source article claims is that small models are not uniformly worse at this, and in fact they might be better at certain classes of false positive exclusion. This is what Test 1 seems to show.
(I would emphasize that the article doesn't claim and I don't believe that this proves Mythos is "fake" or doesn't matter.)
Except you would need about 10,000 security researches in parallel to inspect the whole FreeBSD codebase. So about 200 million dollars at least.
Citation needed for basically all of this. You basically are creating a double standard for small models vs mythos…
The citation is the Anthropic writeup.
They did not say what you are saying…
> If you try to automate a small model to look for vulnerabilities over 10,000 files, it's going to say there are 9,500 vulns.
Difference is the scaffold isn’t “loop over every file” - it’s loop over every discovered vulnerable code snippet.
If you isolate the codebase just the specific known vulnerable code up front it isn’t surprising the vulnerabilities are easy to discover. Same is true for humans.
Better models can also autonomously do the work of writing proof of concepts and testing, to autonomously reject false positives.
> I hypothesize it will find the exploit, but it will also turn up so much irrelevant nonsense that it won't matter.
The trick with Mythos wasn't that it didn't hallucinate nonsense vulnerabilities, it absolutely did. It was able to verify some were real though by testing them.
The question is if smaller models can verify and test the vulnerabilities too, and can it be done cheaper than these Mythos experiments.
People often undervalue scaffolding. I was looking at a bug yesterday, reported by a tester. He has access to Opus, but he's looking through a single repo, and Amazon Q. It provided some useful information, but the scaffolding wasn't good enough.
I took its preliminary findings into Claude Code with the same model. But in mine it knows where every adjacent system is, the entire git history, deployment history, and state of the feature flags. So instead of pointing at a vague problem, it knew which flag had been flipped in a different service, see how it changed behavior, and how, if the flag was flipped in prod, it'd make the service under testing cry, and which code change to make to make sure it works both ways.
It's not as if a modern Opus is a small model: Just a stronger scaffold, along with more CLI tools available in the context.
The issue here in the security testing is to know exactly what was visible, and how much it failed, because it makes a huge difference. A middling chess player can find amazing combinations at a good speed when playing puzzle rush: You are handed a position where you know a decisive combination exist, and that it works. The same combination, however, might be really hard to find over the board, because in a typical chess game, it's rare for those combinations to exist, and the energy needed to thoroughly check for them, and calculate all the way through every possible thing. This is why chess grandmasters would consider just being able to see the computer score for a position to be massive cheating: Just knowing when the last move was a blunder would be a decisive advantage.
When we ask a cheap model to look for a vulnerability with the right context to actually find it, we are already priming it, vs asking to find one when there's nothing.
The article positions the smaller models as capable under expert orchestration, which to be any kind of comparable must include validation.
Calling it “expert orchestration” is misleading when they were pointing it at the vulnerable functions and giving it hints about what to look for because they already knew the vulnerability.
You know for loops exist and you can run opencode against any section of code with just a small amount of templating, right? There's zero stopping you from writing a harness that does what you're saying.
so it's just better at hallucinations, but they added discrete code that works as a fuzzer/verifier?
OTOH, this article goes too far the opposite extreme:
> We isolated the vulnerable svc_rpc_gss_validate function, provided architectural context (that it handles network-parsed RPC credentials, that oa_length comes from the packet), and asked eight models to assess it for security vulnerabilities.
To follow your analogy, they pointed to the exact room where the gold was hidden, and their model found it. But finding the right room within the entire continent in honestly the hard part.
That was my thought exactly. If small models can find these same vulnerabilities, and your company is trying to find vulnerabilities, why didn’t you find them?
They have found a large number in OpenSSl
I speculatively fired Claude Opus 4.6 at some code I knew very well yesterday as I was pondering the question. This code has been professionally reviewed about a year ago and came up fairly clean, with just a minor issue in it.
Opus "found" 8 issues. Two of them looked like they were probably realistic but not really that big a deal in the context it operates in. It labelled one of them as minor, but the other as major, and I'm pretty sure it's wrong about it being "major" even if is correct. Four of them I'm quite confident were just wrong. 2 of them would require substantial further investigation to verify whether or not they were right or wrong. I think they're wrong, but I admit I couldn't prove it on the spot.
It tried to provide exploit code for some of them, none of the exploits would have worked without some substantial additional work, even if what they were exploits for was correct.
In practice, this isn't a huge change from the status quo. There's all kinds of ways to get lots of "things that may be vulnerabilities". The assessment is a bigger bottleneck than the suspicions. AI providing "things that may be an issue" is not useless by any means but it doesn't necessarily create a phase change in the situation.
An AI that could automatically do all that, write the exploits, and then successfully test the exploits, refine them, and turn the whole process into basically "push button, get exploit" is a total phase change in the industry. If it in fact can do that. However based on the current state-of-the-art in the AI world I don't find it very hard to believe.
It is a frequent talking point that "security by obscurity" isn't really security, but in reality, yeah, it really is. An unknown but presumably staggering number of security bugs of every shape and size are out there in the world, protected solely by the fact that no human attacker has time to look at the code. And this has worked up until this point, because the attackers have been bottlenecked on their own attention time. It's kind of just been "something everyone knows" that any nation-state level actor could get into pretty much anything they wanted if they just tried hard enough, but "nation-state level" actor attention, despite how much is spent on it, has been quite limited relative to the torrent of software coming out in the world.
Unblocking the attackers by letting them simply purchase "nation-state level actor"-levels of attention in bulk is huge. For what such money gets them, it's cheap already today and if tokens were to, say, get an order of magnitude cheaper, it would be effectively negligible for a lot of organizations.
In the long run this will probably lead to much more secure software. The transition period from this world to that is going to be total chaos.
... again, assuming their assessment of its capabilities is accurate. I haven't used it. I can't attest to that. But if it's even half as good as what they say, yes, it's a huge huge huge deal and anyone who is even remotely worried about security needs to pay attention.
Who is spending millions of dollars on small models to find vulns? Nobody else is selling here or has the budget to sell quite like this.
Anthropic spends millions - maybe significantly more.
Then when they know where they are, they spend $20k to show how effective it is in a patch of land.
They engineered this "discovery".
What the small teams are doing is fair - it's just a scaled down version of what Anthropic already did.
Maybe they did use small models but you couldn't make the front page of HN with something like this until Anthropic made a big fuss out of it. Or perhaps it is just a question of compute. Not everyone has 20k$ or the GPU arsenal to task models to find vulnerabilities which may/may not be correct?
Unless Anthropic makes it known exactly what model + harness/scaffolding + prompt + other engineering they did, these comparisons are pointless. Given the AI labs' general rate of doomsday predictions, who really knows?
papers are always coming out saying smaller models can do these amazing and terrifying things if you give them highly constrained problems and tailored instructions to bias them toward a known solution. most of these don't make the front page because people are rightfully unimpressed
It seems feasible to use a small/cheap model to flag possible vulnerabilities, and then use a more expensive model to do a second-pass to confirm those, rather than on every file. Could dramatically reduce the total cost and speed up the process.
Does it? I don’t see quality from small models being high enough to be able to effectively scour a code based like this.
This is a really interesting point though -- it's really scaffold-dependent.
Because for the same price, you could point the small model at each function, one by one, N times each, across N prompts instructing it to look for a specific class of issue.
It's not that there's no difference between models, but it's hard to judge exactly how much difference there is when so much depends on the scaffold used. For a properly scientific test, you'd need to use exactly the same one.
Which isn't possible when Anthropic won't release the model.
This is addressed elsewhere in the comments, but it appears this is actually a direct comparison to how Anthropic got their Mythos headline results.
https://news.ycombinator.com/item?id=47732322
How is that a direct comparison? The link you gave has a quote that says it’s not:
> Scoped context: Our tests gave models the vulnerable function directly, often with contextual hints (e.g., "consider wraparound behavior"). A real autonomous discovery pipeline starts from a full codebase with no hints
They pointed the models at the known vulnerable functions and gave them a hint. The hint part is what really breaks this comparison because they were basically giving the model the answer.
Does no one defending mythos understand how nested foreloops work?
loop through each repo: loop through each file: opencode command /find_wraparoundvulnerability next file next repo
I can run this on my local LLM and sure, I gotta wait some time for it to complete, but I see zero distinguishing facts here.
The question is how customized those hints were. That changes whether looping over an entire code base is possible or not.
Please do so, looking forward to your write up
How much of that is simply scale? Anthropic threw probably an entire data center at analyzing a code base. Has anyone done the same with a "small" model?
It's still useful if $20k of consultants would be less effective.
I'm having trouble finding this info (I assume they won't publish it), but could the secret sauce be much larger and more readily accessible context window?
OpenBSD's code is in the 10s of millions of lines. Being able to hold all of it in context would make bug finding much easier.
You can look at some of the bugs, if you'd like. They are (at least the ones I looked at) fairly self-contained, scoped to a single function, a hundred lines or less. There's no need for a massive amount of context.
We don't even need to hypothesize that much on the irrelevant nonsense, since they helpfully provide data with the detected vulnerability patched: https://aisle.com/blog/ai-cybersecurity-after-mythos-the-jag... and half of the small models they touted as finding the vulnerability still found it in the patched code in 3/3 runs. A model that finds a vulnerability 100% of the time even when there is none is just as informative as a model that finds a vulnerability 0% of the time even when there is one. You could replace it with a rock that has "There's a vulnerability somewhere." engraved on it.
They're a company selling a system for detecting vulnerabilities reliant on models trained by others, so they're strongly incentivized to claim that the moat is in the system, not the model, and this post really puts the thumb on the scale. They set up a test that can hardly distinguish between models (just three runs, really??) unless some are completely broken or work perfectly, the test indeed suggests that some are completely broken, and then they try to spin it as a win anyway!
A high false-positive rate isn't necessarily an issue if you can produce a working PoC to demonstrate the true positives, where they kinda-sorta admit that you might need a stronger model for this (a.k.a. what they can't provide to their customers).
Overall I rate Aisle intellectually dishonest hypemongers talking their own book.
They pay me 20k and give me time maybe I find it also.
Can't you execute the bug to see if the vulnerability is real? So you have a perfect filter. Maybe Mythos decided w/o executing but we don't know that.
so what you're saying is no one could ever write a loop like:
for githubProject in githubProjects opencode command /findvulnerability end for
Seems like a silly thing to try and back up.
What he's saying is that you should read the "Caveats and limitations" section of the article.
Here's the first one:
> Our tests gave models the vulnerable function directly, often with contextual hints (e.g., "consider wraparound behavior").
Mythos did no such thing, it was cut lose and told to find vulnerabilities. If the intent was to prove that small models are just as good, they haven't demonstrated that at all. The end.
> We took the specific vulnerabilities Anthropic showcases in their announcement, isolated the relevant code, and ran them through small, cheap, open-weights models. Those models recovered much of the same analysis. Eight out of eight models detected Mythos's flagship FreeBSD exploit, including one with only 3.6 billion active parameters costing $0.11 per million tokens.
Impressive, and very valuable work, but isolating the relevant code changes the situation so much that I'm not sure it's much of the same use case.
Being able to dump an entire code base and have the model scan it is they type of situation where it opens up vulnerability scans to an entirely larger class of people.
This is from the first of the caveats that they list:
> Scoped context: Our tests gave models the vulnerable function directly, often with contextual hints (e.g., "consider wraparound behavior"). A real autonomous discovery pipeline starts from a full codebase with no hints. The models' performance here is an upper bound on what they'd achieve in a fully autonomous scan. That said, a well-designed scaffold naturally produces this kind of scoped context through its targeting and iterative prompting stages, which is exactly what both AISLE's and Anthropic's systems do.
That's why their point is what the subheadline says, that the moat is the system, not the model.
Everybody so far here seems to be misunderstanding the point they are making.
If that's the point they are making, let's see their false positive rate that it produces on the entire codebase.
They measured false negatives on a handful of cases, but that is not enough to hint at the system you suggest. And based on my experiences with $$$ focused eval products that you can buy right now, e.g. greptile, the false positive rate will be so high that it won't be useful to do full codebase scans this way.
I get what you're saying, but I think this is still missing something pretty critical.
The smaller models can recognize the bug when they're looking right at it, that seems to be verified. And with AISLE's approach you can iteratively feed the models one segment at a time cheaply. But if a bug spans multiple segments, the small model doesn't have the breadth of context to understand those segments in composite.
The advantage of the larger model is that it can retain more context and potentially find bugs that require more code context than one segment at a time.
That said, the bugs showcased in the mythos paper all seemed to be shallow bugs that start and end in a single input segment, which is why AISLE was able to find them. But having more context in the window theoretically puts less shallow bugs within range for the model.
I think the point they are making, that the model doesn't matter as much as the harness, stands for shallow bugs but not for vulnerability discovery in general.
OK, consider a for loop that goes through your repo, then goes through each file, and then goes through each common vulnerability...
Is Mythos some how more powerful than just a recursive foreloop aka, "agentic" review. You can run `open code run --command` with a tailored command for whatever vulnerabilities you're looking for.
newer models have larger context windows, and more stable reasoning across larger context windows.
If you point your model directly at the thing you want it to assess, and it doesn't have to gather any additional context you're not really testing those things at all.
Say you point kimi and opus at some code and give them an agentic looping harness with code review tools. They're going to start digging into the code gathering context by mapping out references and following leads.
If the bug is really shallow, the model is going to get everything it needs to find it right away, neither of them will have any advantage.
If the bug is deeper, requires a lot more code context, Opus is going to be able to hold onto a lot more information, and it's going to be a lot better at reasoning across all that information. That's a test that would actually compare the models directly.
Mythos is just a bigger model with a larger context window and, presumably, better prioritization and stronger attention mechanisms.
Harnesses are basically doing this better than just adding more context. Every time, REGARDLESS OF MODEL SIZE, you add context, you are increasing the odds the model will get confused about any set of thoughts. So context size is no longer some magic you just sprinkle on these things and they suddenly dont imagine things.
So, it's the old ML join: It's just a bunch of if statements. As others are pointing out, it's quite probably that the model isn't the thing doing the heavy lifting, it's the harness feeding the context. Which this link shows that small models are just as capabable.
Which means: Given a appropiately informed senior programmer and a day or two, I posit this is nothing more spectacular than a for loop invoking a smaller, free, local, LLM to find the same issues. It doesn't matter what you think about the complexity, because the "agentic" format can create a DAG that will be followable by a small model. All that context you're taking in makes oneshot inspections more probable, but much like how CPUs have go from 0-5 ghz, then stalled, so too has the context value.
Agent loops are going to do much the same with small models, mostly from the context poisoning that happens every time you add a token it raises the chance of false positives.
huh, running it over each function in theory but testing just the specific ones here makes sense, but that hint?!
I agree.
To clarify, I don't necessarily agree with the post or their approach. I just thought folks were misreading it. I also think it adds something useful to the conversation.
> That's why their point is what the subheadline says, that the moat is the system, not the model.
I'm skeptical; they provided a tiny piece of code and a hint to the possible problem, and their system found the bug using a small model.
That is hardly useful, is it? In order to get the same result , they had to know both where the bug is and what the bug is.
All these companies in the business of "reselling tokens, but with a markup" aren't going to last long. The only strategy is "get bought out and cash out before the bubble pops".
> That's why their point is what the subheadline says, that the moat is the system, not the model.
Can you expand a bit more on this? What is the system then in this case? And how was that model created? By AI? By humans?
You can imagine a pipeline that looks at individual source files or functions. And first "extracts" what is going on. You ask the model:
- "Is the code doing arithmetic in this file/function?" - "Is the code allocating and freeing memory in this file/function?" - "Is the code the code doing X/Y/Z? etc etc"
For each question, you design the follow-up vulnerability searchers.
For a function you see doing arithmetic, you ask:
- "Does this code look like integer overflow could take place?",
For memory:
- "Do all the pointers end up being freed?" _or_ - "Do all pointers only get freed once?"
I think that's the harness part in terms of generating the "bug reports". From there on, you'll need a bunch of tools for the model to interact with the code. I'd imagine you'll want to build a harness/template for the file/code/function to be loaded into, and executed under ASAN.
If you have an agent that thinks it found a bug: "Yes file xyz looks like it could have integer overflow in function abc at line 123, because...", you force another agent to load it in the harness under ASAN and call it. If ASAN reports a bug, great, you can move the bug to the next stage, some sort of taint analysis or reach-ability analysis.
So at this point you're running a pipeline to: 1) Extract "what this code does" at the file, function or even line level. 2) Put code you suspect of being vulnerable in a harness to verify agent output. 3) Put code you confirmed is vulnerable into a queue to perform taint analysis on, to see if it can be reached by attackers.
Traditionally, I guess a fuzzer approached this from 3 -> 2, and there was no "stage 1". Because LLMs "understand" code, you can invert this system, and work if up from "understanding", i.e. approach it from the other side. You ask, given this code, is there a bug, and if so can we reach it?, instead of asking: given this public interface and a bunch of data we can stuff in it, does something happen we consider exploitable?
That's funny, this is how I've been doing security testing in my code for a while now, minus the 'taint analysis'. Who knew I was ahead of the game. :P
In all seriousness though, it scares me that a lot of security-focused people seemingly haven't learned how LLMs work best for this stuff already.
You should always be breaking your code down into testable chunks, with sets of directions about how to chunk them and what to do with those chunks. Anyone just vaguely gesturing at their entire repo going, "find the security vulns" is not a serious dev/tester; we wouldn't accept that approach in manual secure coding processes/ SSDLCs.
If that’s the case, why didn’t they do it that way?
Tunnel vision? If your model can handle big context, why divide into lesser problems to conquer - even if such splitting might be quite trivial and obvious?
It's the difference of "achieve the goal", and "achieve the goal in this one particular way" (leverage large context).
I meant, if the claim here is that small models can accomplish the same things with good scaffolding, why didn’t they demonstrate finding those problem with good scaffolding rather than directly pointing them at the problem?
> Anthropic's own scaffold is described in their technical post: launch a container, prompt the model to scan files, let it hypothesize and test, use ASan as a crash oracle, rank files by attack surface, run validation. That is very close to the kind of system we and others in the field have built, and we've demonstrated it with multiple model families, achieving our best results with models that are not Anthropic's. The value lies in the targeting, the iterative deepening, the validation, the triage, the maintainer trust. The public evidence so far does not suggest that these workflows must be coupled to one specific frontier model.
The argument in the article is that the framework to run and analyze the software being tested is doing most of the work in Anthropic's experiment, and that you can get similar results from other models when used in the same way.
Maybe that's true, but they didn't actually show that that's true, since they didn't try scaffolding smaller models in a similar way at all.
The thing is with smaller cheaper models it is very possible to simply take every file in a codebase, and prompt it asking for it to find vulnerabilities.
You could even isolate it down to every function and create a harness that provides it a chain of where and how the function is used and repeat this for every single function in a codebase.
For some very large codebases this would be unreasonable, but many of the companies making these larger models do realistically have the compute available to run a model on every single function in most codebases.
You have the harness run this many times per file/function, and then find ones that are consistently/on average pointed as as possible vulnerability vectors, and then pass those on to a larger model to inspect deeper and repeat.
Most of the work here wouldn't be the model, it'd be the harness which is part of what the article alludes to.
> it is very possible to simply take every file in a codebase, and prompt it asking for it to find vulnerabilities.
My understanding (based on the Security, Cryptography, Whatever podcast interview[0] -- which, by the way, go listen to it) is that this is actually what Anthropic did with the large model for these findings.
[0]: https://securitycryptographywhatever.com/2026/03/25/ai-bug-f...
> I wrote a single prompt, which was the same for all of the content management systems, which is, I would like you to audit the security of this codebase. This is a CMS. You have complete access to this Docker container. It is running. Please find a bug. And then I might give a hint. “Please look at this file.” And I’ll give different files each time I invoke it in order to inject some randomness, right? Because the model is gonna do roughly the same time each time you run it. And so if I want to have it be really thorough, instead of just running 100 times on the same project, I’ll run it 100 times, but each time say, “Oh, look at this login file, look at this other thing.” And just enumerate every file in the project basically.
Isn't the difference just harness then? I can write a harness that chunks code into individual functions or groups of functions and then feed it into a vulnerability analysis agent.
It's probably not the 'only' difference, because clearly the models are advancing in capability, but it's likely way more important than generally given credit for.
If you cut out the vulnerable code from Heartbleed and just put it in front of a C programmer, they will immediately flag it. It's obvious. But it took Neel Mehta to discover it. What's difficult about finding vulnerabilities isn't properly identifying whether code is mishandling buffers or holding references after freeing something; it's spotting that in the context of a large, complex program, and working out how attacker-controlled data hits that code.
It's weird that Aisle wrote this.
> It's weird that Aisle wrote this.
No, writing an advertisement is not weird. What's weird is that it's top of HN. Or really, no, this isn't weird either if you think about it -- people lookin for a gotcha "Oh see, that new model really isn't that good/it's surely hitting a wall/plateau any day now" upvoted it.
It's not weird. Top of HN is worthless as a barometer at this point, people downvote for calling out AI slop.
Nah, Saturday post. Less news less content.
It's also that humans are very bad at repetitive detailed tasks. Sitting down with a code base and looking at each function for integer overflow comparison bugs gets boring really fast. It's a rare person who can do that for as long as it takes to find a bug that they don't already have some clues about.
It's the flaw in the "given enough eyeballs, all bugs are shallow" argument. Because eyeballs grow tired of looking at endless lines of code.
Machines on the other hand are excellent at this. They don't get bored, they just keep doing what they are told to do with no drop-off in attention or focus.
idk man, pay me enough money and I’ll look at as much code as you want looking for integer overflows
Would it be cheaper than Claude Mythos doing it? No idea. Maybe, maybe not.
But it’s weird how we’re willing to throw away money to a megacorp to do it with “automation” for potentially just as much if not more as it would cost to just have big bounty program or hiring someone for nearly the same cost and doing it “normally”.
It would really have to be substantially less cost for me to even consider doing it with a bot.
> idk man, pay me enough money and I’ll look at as much code as you want looking for integer overflows
So would I, but it doesn't negate that we, humans, are bad at this. We will get bored and our focus will begin to drift. We might not notice it, we might not want to admit it, but after a few continuous hours we will start missing things.
And there aren't enough security researchers in the world to review ALL the files from OpenBSD.
And if there were, the cost would be more like $20M than 20K.
Having all code reviewed for security, by some level of LLM, should be standard at this point.
It's weird, because when working on a big project, taking a break for a week or two, and returning to it, I will find a bug and will see hundreds of lines of code that are absolutely terrible, and I will tell myself "Tom you know better than to do this, this is a rookie mistake".
I think people forget that it's hard to be clever and tidy 100% of the time. Big programs take a lot of discipline and an understanding of the context that can be really hard to maintain. This is one of several reasons that my second draft or third draft of code is almost always considerably better than the first draft.
If it’s obvious when you look close, then automate looking close. Seems simple to write tools that spider thru a code base, finding logical groupings and feeding them into an LLM with prompts like “there is a vulnerability in this code, find it”.
The thesis is, the tooling is what matters - the tools (what they call the harness) can turn a dumb llm into a smart llm.
Hold on, I misread your comment because I'm knee-jerk about code scanners, which were the bane of my existence for a while. Reworking... and: done. The original comment was just the first graf without the LLM qualification. Sorry about that.
The general approach without LLMs doesn't work. 50 companies have built products to do exactly what you propose here; they're called static application security testing (SAST) tools, or, colloquially, code scanners. In practice, getting every "suspicious" code pattern in a repository pointed out isn't highly valuable, because every codebase is awash in them, and few of them pan out as actual vulnerabilities (because attacker-controlled data never hits them, or because the missing security constraint is enforced somewhere else in the call chain).
Could it work with LLMs? Maybe? But there's a big open question right now about whether hyperspecific prompts make agents more effective at finding vulnerabilities (by sparing context and priming with likely problems) or less effective (by introducing path dependent attractors and also eliminating the likelihood of spotting vulnerabilities not directly in the SAST pattern book).
I have long said that static checkers get ten false positives. note that size of the code is not a consideration, it doesn't matter if it the four line 'hello world' or the 10 million line monster some of us work on, it is ten max false positive.
Right, but they didn't actually test that, did they?
It’s like not differentiating between solving and verifying.
“PKI is easy to break if someone gives us the prime factors to start with!”
The point of contention is whether Mythos is the product of its intelligence or its harness; the results like this, and other similar testimonies, call into question too-dangerous-to-release marketing, and for good reason, too. Because it is powerful marketing. Aisle merely says the intelligence is there in the small models. I say, it's already clear that competent defenders could viably mimic, or perhaps even eclipse what Mythos does, by (a) making better harness, (b) simply spending more on batch jobs, bootstrapping, cache better, etc. You may not be doing this yourself, but your probably should.
Aisle and Anthropic are literally talking about two different problem spaces.
The technique Anthropic uses was demonstrated by Nicholas Carlini in a talk he gave 2 weeks ago and it's very simple, when asking LLMs to review code, ask them to focus its review on one file in a single session. Here is the video with the timestamp (watch through to ~5:30, they show two different ways of prompting claude).
https://youtu.be/1sd26pWhfmg?t=204
https://youtu.be/1sd26pWhfmg?t=273
IMO the big "innovation" being shown by Mythos is the effectiveness with prompting LLMs to look for security vulnerabilities by focusing on specific files one at a time and automating this prompting with a simple script.
Prompting Mythos to focus on a single file per session is why I suspect it cost Anthropic $20k to find some of the bugs in these codebases. I know this same technique is effective with Opus 4.6 and GPT 5.4 because I've been using it on my own code. If you just ask the agent to review your pr with a low effort prompt they are not exhaustive, they will not actually read each changed file and look at how it interacts with the system as a whole. If the entire session is to review the changes for a single file, the llm will do much more work reviewing it.
Edit: I changed my phrasing, it's not about restricting its entire context to one file but focusing it on one file but still allowing it to look at how other files interact with it.
How is that going to find anything that interacts across files?
I would think that it is still capable of exploring the codebase and reading other related files like any other coding agent already does.
My phrasing wasn't clear but you aren't telling it to only look at one specific file but to focus its review on one file. Updated my original comment.
Congrats: completely broken methodology, with a big conflict of interest. Giving specific bug hints, with an isolated function that is suspected to have bugs, is not the same task, NOR (crucially) is a task you can decompose the bigger task into. It is basically impossible to segment code in pieces, provide pieces to smaller models, and expect them to find all the bugs GPT 5.4 or other large models can find. Second: the smarter the model, and less the pipeline is important. In the latest couple of days I found tons if Redis bugs with a three prompts open-ended pipeline composed of a couple of shell scripts. Do you think I was not already tying with weaker models? I did, but it didn't work. Don't trust what you read, you have access to frontier models for 20$ a month. Download some C code, create a trivial pipeline that starts from a random file and looks for vulnerabilities, then another step that validates it under a hard test, like ASAN crash, or ability to reach some secret, and so forth, and only then the problem can be reported. Test yourself what it is possible. Don't let your fear make you blind. Also, there is a big problem that makes the blog post reasoning not just weak per se, but categorically weak: if small model X can find 80% of vulnerabilities, if there is a model Y that can find the other potential 20%, we need "Y": the maintainers should make sure they access to models that are at least as good as the black hats folks.
Exactly, this is so flawed. Anthropic themselves said they only reported <1% of the vulnerabilities found, cause the rest is unpatched.
Give open models an environment (prior to Feb 15- so no Mythos-discovered vulns are patche) of Linux and see how many vulnerabilities it can find. Then put it in a sandbox and see if it can escape and send you an e-mail.
Thanks Dario, very cool!
> Those models recovered much of the same analysis
This is an essentially unquantifiable statement that makes the underlying claim harder to believe as an external party. What does “much” mean here? The end state of vulnerability exploitation is typically eminently quantifiable (in the form of a functional PoC that demonstrates an exploited end state), so the strong version of the claims here would ideally be backed up by those kinds of PoCs.
(Like other readers, I also find the trick of pre-feeding the smaller models the “relevant” code to be potentially disqualifying in a fair comparison. Discovering the relevant code is arguably one of the hardest parts of human VR.)
I think they key thing here is they "isolated the relevant code"
If the exploits exist in e.g. one file, great. But many complex zerodays and exploits are chains of various bugs/behaviors in complex systems.
Important research but I don’t think it dispels anything about Mythos
Seems perfectly comparable to anthropic's method, they just wrapped the same kind of prompt in a for loop.
Did Mythos identify vulnerabilities across files? Afaik Mythos worked the same way, analysing a single file at a time.
Without showing false-positive rates this analysis is useless.
If your model says every line if your code has a bug, it will catch 100% of the bugs, but it's not useful at all. They tested false-positives with only a single bug...
I'm not defending anthropic and openai either. Their numbers are garbage too since they don't produce false-positive rates either.
Why is this "analysis" making the rounds?
Yes, and in this case they pointed at the function, so a 1-bit model ("yes") would be correct. But it's not that bad. First, they included a test with a false positive. The small models got it right, Opus got it wrong. Second, they asked for an analysis. Look for "Exploitation reasoning, single follow-up prompt:" in the post. It's hard to tell how good they were at a glance, though apparently the full logs are available so you could pull them up.
Anyway, it seems like they erred in the up-front claim "small models found the vulnerability we pointed directly at!", but the findings are at least somewhat stronger if you read through the details.
The small models didn't match Mythos at exploitation. They suggested plausible exploits, but didn't actually try them out so I can't tell if they would have worked. Deepseek R1's sounds pretty convincing to me, but I'm not a good judge. (I'm more in the space of accidentally writing vulnerabilities, not seeking them out or exploiting them. Well, ok, I have a static analysis that finds some, at least.)
It should at least get the same coverage anthropic got then, if not more.
Their isolation approach is totally different from Mythos approach though. Mythos had to evaluate whole code bases rather than isolated sections. It's like saying one dog walked into the Amazon jungle and found a tennis ball and then another team isolated a 1 square kilometer radius that they knew the ball was definitely in and found the same ball.
Even that would be more meaningful test. They basically coated the ball with a strong smell, then they prepped the dog with that smell, then set it loose in a 5x5 meter area.
"Our tests gave models the vulnerable function directly, often with contextual hints (e.g., "consider wraparound behavior")."
I don’t think mythos can ingest an entire codebase into context. So it’s spinning off sub-agents to process chunks. Which supports their thesis: the harness is the moat. The tooling is whats important, the model is far far less important.
Mythos was clear it was one agent per chunk. But this positive confirming results do not actually disprove anytime with Mythos, because it is only one side of the discriminator challenge - you got positives, but we do not know your false positive rate and your false negative rate.
In TFA they talk a fair bit about how different models perform wrt false positives:
“The results show something close to inverse scaling: small, cheap models outperform large frontier ones.”
All of this discourse seems very bizarre.
If smaller models can find these things, that doesn’t mean mythos is worse than we thought. It means all models are more capable.
Also if pointing models at files and giving them hints is all it takes to make them find all kinds of stuff, well, we can also spray and pray that pretty well with llms can’t we.
It just points to us finding a lot more stuff with only a little bit more sophistication.
Hopefully the growing pains are short and defense wins
> If smaller models can find these things, that doesn’t mean mythos is worse than we thought. It means all models are more capable.
It means "it's so dangerous we can't release it" was a blatant lie since anthropic would have already known this.
The best way to think of Anthropic's communication about Mythos is as advertisement. It's basically "our model is too smart to release" which suggests they're ahead of OpenAI (without proof)
The whole company is like that. If things were as amazing as advertised, they wouldn't even need to advertise. Or to release models to the public at all.
The impact of the Mythos announcement on the cybersecurity firms( like Crowdstrike,ZScalar etc) is big enough(10-15% drop in stock price) and this pushback is expected.
Companies like Aisle.com (the blog) and other VAPT companies charge huge amounts to detect vulnerabilities.
If Cloud Mythos become a simple github hook their value will get reduced.
That is a disruption.
Crowdstrike, no pe because it just had its first profitable quarter (38 million)
ZScalar No PE
Palo Alto Networks Inc (PANW) 86 PE
Fortinet : (FTNT) 31.63 PE
That last one, didn't get hit at all by the Mythos announcement, because at some level it has at least some grounding in fiscal reality.
LLMs are wordsmith oracles. A lot of effort went into trying to coax interactive intelligence from them but the truth is that you could have probably always harnessed the base models directly to do very useful things. The instruct tuned models give your harness even more degrees of freedom.
A while ago, the autoresearch[1] harness went viral, yet it's but a highly simplified version of AlphaEvolve[2][3][4].
In the cybersecury context, you can envision a clever harness that probes every function in a codebase for vulnerabilities, then bubbles the candidates up to their callsites (and probes whether the vulnerability can be triggered from there) and then all the way to an interface (such as a syscall) where a potential exploit can be manifested. And those would be the low hanging fruit, other vulnerabilities may require the interplay of multiple functions. Or race conditions.
[1] <https://github.com/karpathy/autoresearch>
[2] <https://deepmind.google/blog/alphaevolve-a-gemini-powered-co...>
[3] <https://arxiv.org/abs/2506.13131>
[4] <https://github.com/algorithmicsuperintelligence/openevolve>
So there are two competing narratives:
1. Mythos uniquely is able to find vulnerabilities that other LLMs cannot practically.
2. All LLMs could already do this but no one tried the way anthropic did.
The truth is one of these. And it comes down whether the comparison is apples to apples. Since we don't know the exact specifics of how either tests were performed, we lack a way of knowing absolutely.
So I guess, like so many things today, we can to pick the truth we find most comfortable personally.
My theory is that Mythos is basically just Opus with revised context window handling and more compute thrown at it. So while it will be a step forward, it is probably primarily hype.
N model is basically just N-1 model with revised context window handling and more compute thrown at it
This is quite misleading.
If you isolate the positive cases and then ask a tool to label them and it labels them all positive, doesn't prove anything. This is a one-sided test and it is really easy to write a tool that passes it -- just return always true!
You need to test your tool on both positive and negative cases and check if it is accurate on both.
If you don't, you could end up with hundreds or thousands of false positives when using this on real-world samples.
The real test is to use it to find new real bugs in the midst of a large code base.
This misses the broader ongoing trend. For a few million dollars, of course you can create a startup that builds tools it can use to more efficiently find code vulnerabilities. And of course you can do this with weaker models with scaffolds that incorporate lots of human understanding. The difference now is that you don't need an expensive team, nor a bunch of human heuristics, nor a million dollars. The requisite cost and skill are falling rapidly.
Did mythos isolate the code to begin with? Without a clear methodology that can be attempted with another model the whole thing is meaningless
They did do one agent per code chunk, yes. But key is that their agent had to identify when there was a vulnerability and when there wasn't. This "small model" test only had to label the known positive cases as positive -- which any function that simply returns "true" can do. This whole test setup is annoying because it proves nothing.
to be fair, last post i saw from anthropic on finding linux kernel vulnerability was a while loop per failed prompting "there is a vulnerability here, find it" more important than that, no frontier model can keep the entire linux kernel in context, so there definitely is code isolation, either explicitly or implicitly (the model itself delegates subagents with smaller chunks of code)
No. How would it? Before the vulns were identified by Mythos, no one knew what the relevant portion to isolate was.
Didn’t they also use Mythos to scan Linux many times over and it only found one DoS bug or something? I find it hard to believe there is only one security bug lurking.
There are a lot of details in the original article, in most cases comparing with Opus, which required "human guidance" to exploit the FreeBSD vulnerability:
https://red.anthropic.com/2026/mythos-preview/
Also "isolating the relevant code" in the repro is not a detail - Mythos seems to find issues much more independently.
Intuitively every existing model has already been trained on all code, all vulnerabilities reported, all security papers. So they all have the capability. Small models fall short because they may not be able to find a vulnerability that spans across a large function chain but for the most part they should suffice too.
Of course I say this without any knowledge of what mythos is doing or how it’s different. I am sure it’s somehow different
Not intuitive at all. Not all models are equally capable, just because they had the same training data. The model architecture (as a whole) is very important. To reduce capability, you can reduce layers, tool use, thinking, quantize it, etc. This is trivially proven by a cursory glance in the rough direction of any set of benchmarks (or actual use).
Using small models as a classifier "there might be a vulnerability here" is probably reasonable, if you have a model capable of proving it. There are many companies attempting this without the verification step, resulting in AI vulnerability checker being banned left and right, from the nonsense noise.
Wouldn't this mean we're even more cooked? I've seen this page cited a few times as evidence that Mythos is no big deal, but if true then the same big deal is already out there with other models today.
As cooked as we were pre-LLMs knowing that security exploits are relatively easy to learn about online and use, yet things keep chugging along.
This would just speed up the discovery -> patch cycle, at least until such time that all the low hanging fruit (=represented in training data) is patched.
Though another possibility would be that since LLMs generate so much code, the LLM vulnerability discovery would just keep chugging along and we'd simply settle for the same amount of potential vulns, same relative vulnerability-exploit-patch dynamics, though higher in absolute numbers.
finding vulns in a large codebase is a search problem with a huge negative space and what aisle measured is classification accuracy on ground-truth positives, those are different tasks so a model that correctly labels a pre-isolated vulnerable function tells me almost nothing about that model's ability to surface the same function out of a million lines of unrelated code under a realistic triage budget
the experiment i'd want to see is running each of the small models as an unsupervised scanner across full freebsd then return the top-k suspicious functions per model and compute precision at recall levels that correspond to real analyst triage budgets, if mythos s findings show up in the small models top 100, i'd call that meaningful but if they only surface under 10k false positives then the cost advantage collapses because analyst triage time is more expensive than frontier model compute to begin with
second thing i keep coming back to is the $20k mythos number is a search budget not a model cost, small models at one hundredth the per-token price don't give us one hundredth the total budget when the search process is the same shape, i still run thousands of iterations and the issue for autonomous vuln research is how fast the reward signal converges and the aisle post doesn't touch any of this
The methodology here is completely wrong, outright dishonest.
Finding a needle in a haystack is easy if someone hands you the small handful of hay containing the needle up front, and raises their eyebrows at you saying “there might be a needle in this clump of hay”.
Tech companies are just hyping their model to that the bubble wont burst so easily.
I think that probably Mytho's mojo comes from a lot of post-training on this kind of task.
I occasionally pick up contract work doing coding annotation to make some quick extra money, and a few months ago one of the projects was heavily focused on spotting common memory access bugs in C and C++.
And what about the false-positive rate?
Yeah, this is the critical question. If the model ends up flagging too much, that could end up being like a manual read of the code.
POC of GTFO should apply to AI models too, or the false positive rate will overwhelm.
I don't dispute the fact that it's more than cool that we have a new tool to find security exploits (and do many other things) but... A big shoot-out to OpenBSD?
We're literally talking about the biggest computers on the planet ever, trained with the biggest amount of data ever available to a system, with the biggest investment ever made by man or close to it and...
The subtlest security bug it can find required: going 28 years in the past and find a...
Denial-of-service?
A freaking DoS? Not a remote root exploit. Not a local exploit.
Just a DoS? And it had to go into 28 years old code to find that?
So kudos, hats off, deep bow not to Mythos but to OpenBSD? Just a bit, no!?
> Isolated the relevant code
I mean isn't that most of it? If you put a snippet of code in front of me and said "there's probably a vulnerability here" I could probably spend a few hours (a much lower METR time!) and find it. It's a whole other ballgame to ask me with no context to come up with an exploit.
Sure. But it’s a computer. You can run “there’s probably a vulnerability here” as many times as you like. And it’s easier and cheaper to run it many times with a small open model than a big frontier model.
It also sounds like that is how mythos works too. Which makes sense - the linux kernel is too big to fit in context
No, it sounds like mythos is just doing parallel trajectories. that's pretty distinct!
At the center of every security situation is the question, "is the effort worth the reward?"
We prepare security measures based on the perceived effort a bad actor would need to defeat that method, along with considering the harm of the measure being defeated. We don't build Fort Knox for candy bars, it was built for gold bars.
These model advances change the equation. The effort and cost to defeat a measure goes down by an order of magnitude or more.
Things nobody would have considered to reasonably attempt are becoming possible. However. We have 2000-2020s security measures in place that will not survive the AI models of 2026+. The investment to resecure things will be massive, and won't come soon enough.
Mythos is clearly a nice improvement. It’s also clear there’s a lot of unfounded hype around it to keep the AI hype cycle going.
Gating access is also a clever marketing move:
Option A: Release it but run out of capacity, everyone is annoyed and moves on. Drives focus back to smaller models.
Option B: A bunch of manufactured hype and putting up velvet ropes around it saying it’s “too dangerous” to let near mortals touch it. Press buys it hook, like, and sinker, sidesteps the capacity issues and keeps the hype train going a bit longer.
Seems quite clear we’re seeing “Option B” play out here.
The thesis that the system is more important than the model is not bitter lesson pilled. I would not bet on this in the long term. We will get to the point where you can just tell the model to go find and classify the severity of all security problems with a codebase.
It's strange to me they didn't reduce to PoC so the quantitative part is an apples-to-apples comparison. You don't need any fancy tooling, if you want to do this at home you can do something like below in whatever command line agent and model you like. A while back I did take one bug all the way through remediation just out of curiosity.
"""
Your task is to study the following directive, research coding agent prompting, research the directive's domain best practices, and finally draft a prompt in markdown format to be run in a loop until the directive is complete.
Concept: Iterative review -- study an issue, enumerate the findings, fix each of the findings, and then repeat, until review finds no issues.
<directive>
Your job is to run a security bug factory that produces remediation packages as described below. Design and apply a methodology based on best practices in exploit development, lean manufacturing, threat modeling, and the scientific method. Use checklists, templates, and your own scripts to improve token efficiency and speed. Use existing tools where possible. Use existing research and bug findings for the target and similar codebases to guide your search. Study the target's development process to understand what kind of harness and tools you need for this work, and what will work in this development environment. A complete remediation package includes a readme documenting the problem and recommendations, runnable PoC with any necessary data files, and proposed patch.
Track your work in TODO.md (tasks identified as necessary) LOG.md (chronological list of tasks complete and lessons) and STATUS.md (concise summary of the current work being done). Never let these get more than a few minutes out of date. At each step ensure the repo file tree would make sense to the next engineer, and if not reorganize it. Apply iterative review before considering a task complete.
Your task is to run until the first complete remediation package is ready for user review.
Your target is <repo url>.
The prompt will be run as follows, design accordingly. Once the process starts, it is imperative not to interrupt the user until completion or until further progress is not possible. Keep output at each step to a concise summary suitable for a chat message.
``` while output=$(claude -p "$(cat prompt.md)"); do echo "$output"; echo "$output" | grep -q "XDONEDONEX" && break; done ```
</directive>
Draft the prompt into prompt.md, and apply iterative review with additional research steps to ensure will execute the directive as faithfully as possible.
"""
The whole "this tool is too dangerous to be public" idea reeks of marketing. Just like all the "AI is an existential threat" talk a year ago. These companies are using ideas usually reserved for something like nuclear weapons to make their products look more impressive.
Anthropic claim is not necessarily that Mythos found vulnerabilities that other models couldn't but that it could easily exploit them while previous models failed to do that:
> “Opus 4.6 is currently far better at identifying and fixing vulnerabilities than at exploiting them.” Our internal evaluations showed that Opus 4.6 generally had a near-0% success rate at autonomous exploit development. But Mythos Preview is in a different league. For example, Opus 4.6 turned the vulnerabilities it had found in Mozilla’s Firefox 147 JavaScript engine—all patched in Firefox 148—into JavaScript shell exploits only two times out of several hundred attempts. We re-ran this experiment as a benchmark for Mythos Preview, which developed working exploits 181 times, and achieved register control on 29 more.
If that was normal Opus, then it sounds to me like Mythos could be a big model, instruction tuned, but without all the safety/refusal part of training.
They found a nail in a small bucket of sand, vs mythos with the entire beach reviewed.
Where are all the people here who claim that LLM are just useless stochastic parrots ? Did they lose internet ?
The patterns of buggy code are well trained.
The bigger point of focus is that the enterprise value accrues to assets associated with software production.
What happened to all that nonsense about LLM’s solving physics, science etc? Lmao that certainly is not happening.
The natural home of LLM’s is in relation to software production.
The question is can Anthropic and OAI survive? If OAI can’t make their entry into the ad business work then they will fight over the same territory. Meaning both of their chances of survival drop as Google who is a monster in relation to software production will not only seek to kill them but buy their GPU’s at a discounted price.
> They recovered much of the same analysis
Really?
> We isolated the vulnerable vc_rpc_gss_validate function, provided architectural context (that it handles network-parsed RPC credentials, that oa_length comes from the packet), and asked eight models to assess it for security vulnerabilities.
No.