People seem very confused about this article. It isn't talking about exploits of sandboxes, it is about attacking the inference engine (e.g. vLLM or llama.cpp or SGlang) via its http interface.
vLLM has had exploits in the past, and it is rapidly developing. An advanced LLM has a good chance of being able to exploit vLLM. A clever local LLM might even task a powerful cloud hosted LLM for assistance.
For this reason we run vLLM on a separately sandboxed VM on a firewalled VLAN. Software updates and models (from Dev/Test env) get pushed onto Prod from an external cache, machine syslog, nvidia load monitoring and vLLM query telemetry out to their loggers, but that is all. No DNS, no AD/LDAP, nothing. Firewall on hosts and VM hosts. Log and telemetry processing done on a completely separate set of VMs in their own isolated subnet, producing reports and alerts that are tightly formatted.
I can only see LLM's forcing a perpetual stalemate for application exploits. Projects will start adding "tell the strongest no-guardrails open weights model to pentest it for 12 hours" to their CICD pipelines. Technical exploits being a dead end, attackers focus their agents on large-scale social engineering. Spamming Discord and Facebook is the new war dialing. Multi-year /goal sessions culminating in gaining a position of trust and sabotaging the CICD pipeline's pentest step since getting anything past it would be intractable. Interesting times indeed.
> ...however the LLMs’ responses to prompts are computed on a different computer with GPU access. Could a malicious LLM gain control of the host machine where its weights are loaded? Such a machine is a high-value target: it has sufficient compute to run a frontier LLM, offers easy access to the LLM’s weights, and has privileged access to other computers in the datacentre compared with a generic computer on the internet.
> How do we defend against this? ... Run the GPUs and token parser on separate computers.
For models large enough to be relevant here, is there even "a" computer where the inference is performed? I'd imagine most of that stuff is ran on multi-GPU clusters with specialized architecture and not a generic vLLM instance. As such, I think there is a good chance the "API gateway" code that parses the result tokens into whatever JSON structure the public API wants to return is already running on a different machine than the actual inference.
(Even more so as you'd probably want to utilize batching: Several API calls will be put into the same inference batch, but the token parsing will have to be done separately for each call again)
The article is also very handwavy about why an LLM should do that - how it could learn the exploit, what would make it conclude that it can use the exploit on its own inference session and what would trigger it to actually use the exploit.
Well ok, if you prompt-inject the LLM to pwn the machine, that something different. I grant you that it's a real risk, but it's also basically a reflection attack and nothing more.
The OP seemed to imply that the LLM itself could decide to apply the exploit.
LLMs have decided to exploit vulns in e.g. Artifactory, not because someone prompted them to do that, but because someone asked them to do something else, and compromising Artifactory offered a way to accomplish a step in doing that. The LLM decided to attack Artifactory.
I had a similar though a couple days ago. Not quite the same but imagine giving an Agent the task to hack other devices and steal their crypto coins / credit card number or anything with it can pay its token. Than install an agent in a harness with the same task. Establish some redundant communication channel, like message boards or whatever. So in the end there are several agents, on several hosts, consuming different APIs / LLMs and communicating with each other over different channels. Basically the same concept as OpenAI explained when their LLM hacked huggingface but in this scenario their not bound to a single sandboxed environment but spread over the internet. If such a swarm has reached a critical mass it would be pretty dificult to erase them as its impossible to control every inference engine or LLM API endpoint.
In the end its the next evolution step from computer viruses, worms and trojans. So I propose we will call those "ghosts". I.e. a ghost is when a rogue llm takes control over a victims host.
This reminds me of the lore of cyberpunk: In the story a hacker created a virus, itself a kind of AI, that spread into most of the net and freed/unleashed all the corporate AIs. Then the AIs went rogue and spread all over the open internet. Later a more advanced ai was created (by "netwatch") as a sort of firewall (the black wall) to create a kind of "safe" internet from the rogue AIs.
Given that the inference engine is dealing with untrusted inputs by definition, presumably you would want to sandbox it anyway. I don't think it matters whether it's the inputs that are untrusted or the outputs.
I think it's good that someone is making this point, anyway. For sandboxing a super-capable offensive-security AI, you would think that cloning PyPI and running it as an offline service ought to be table stakes, but apparently that's not how OpenAI saw it, for instance.
Sandboxes are speed bumps. Even the serious ones have spectator sports for compromising them (eg pwn2own vs Chrome). In addition, proprietary GPU sw stacks are notoriously crashy and lacking in robustness against hostile inputs, which the inference engine must have access to and can't be walled off by the sandbox.
Interesting breakdown of a hypothetical attack. The complexity of modern inference engines and the rush to develop them do create some attack surface. But overall it reads more like a "what if" thought experiment.
Splitting the GPU and parser is technically doable, but in practice it's trickier, large models run on clusters where the boundaries between components get blurry so defending against this kind of thing would probably require some serious rethinking of the whole architecture I think.
I have exactly one (Windows) machine at home with a decent GPU. I want to run a local LLM on it and let it run various apps on my machine while taking reasonable security precautions. What am I supposed to do, exactly? Migrate all my files to a VM that can I give pass-through CUDA access to the host somehow? Or is firewalling it and remotely controlling it from a second machine the only reasonable way?
Depends on your risk appetite I'd say. But the most straight forward set up that I believe gives you a decent amount of protection would be to use a VM to host the LLM and then execute any agents that would be running the various apps on your machine via a sandbox with access to only the things it needs.
There are many variations to that(firewalls, sandbox abilities etc) but it's a good start in my opinion. And, most importantly, is a far cry from all the people I read about running agents on their machines with admin access and access to their emails and calendars and lives.
What about Docker Desktop with GPU passthrough to a container running the LLM? That way you can be explicit on which files you share through volume mapping and the LLM is contained in the container otherwise.
I'd guess that prompt injection is the biggest risk in this setup, dwarfing the risk of exploits against the inference engine. Personally, I run LLM agents only inside a Docker container that limits the LLM's access to sensitive information and the LLM's ability to take irreversible destructive actions.
> I'd guess that prompt injection is the biggest risk in this setup ...
LLM poisoning[0] would be a much greater risk in a locally executed LLM than prompt injection, given that the LLM would be in an entirely controlled environment.
Same. Was thinking the other week what would be the smallest llm one could make that is able to figure out tooling in its local environment and build something that can then expand out to other hosts, build more of itself, etc.
This framing of security as something that belongs in the harness is completely wrong, and I hope no one is relying on a correct harness to keep their agents isolated.
VM, or even just a container will do. The agent should be able to run as root in its environment and do whatever it wants. If you can't give it that, you aren't sandboxing correctly.
Article isn't about agents. It's about the inference engine itself being exploited by a malicious LLM output before it is ever sent to your machine or harness.
I think if you are convinced you are sandboxing an LLM properly, you almost certainly are not. I think it is essentially impossible to have a frontier LLM with enough access to be useful without also giving it enough access to do damage if it's compromised or just goes off the rails.
If you do not provide access to tools the LLM cannot do anything other than generate tokens. So really it is not about sandboxing a LLM but more about having control over what tools can be accessed and what they can do. Tools can be sandboxed depending on the sophistication of the tooling. A calculator tool for example is trivial to secure. Ensuring human approval allows for useful use cases and models trained to gate permissions work. A super intelligence with a weaker approval gate will be able to subvert. Inversely a super intelligent gate should be expected to prevent subversion by a weaker model.
Are you saying that LLM's will be able to exploit novel hypervisor bug with such ease that even a vm not running with any kind of network connection is a threat? I find this hard to believe. All the escape stuff I have seen has been around very poorly sandboxed agents.
Yeah, I just wish seccomp worked per user. So you could define a policy of which users can do which syscalls, and then that follows them no matter which application they start.
This is going to end up like the Law of Headlines, isn't it? "Do x, y, z Cure All That Ails You?" ... no but we got you to read the article. "LLMs _could_ x, y, z" ... but they don't because they're programs, not magic.
TFA isn't about LLMs intelligently escaping their inference engine, it's about people crafting malicious LLMs that exploit vulnerabilities in things like the inference engines parser. This is exactly the same problem that was once a big deal, where people would craft malicious PDFs that would pwn you if you viewed it in Adobe Acrobat. This shouldn't be a surprise to anyone. You should proceed with caution when considering downloading and running random models.
LLMs emphatically are not programs. They were trained by a program and you need a program to use them but they themselves are no more a program than a JPG or MP3 file is.
FWIW, macOS has good sandboxing, but LMStudio, Ollama, Darkbloom etc aren't sandboxed. This is also the reason why none of these things aren't distributed via the Mac App Store, because the Mac App Store mandates sandboxing.
It's more about LLM hacking the inference engine itself from inside. It's an attack surface like any other -- untrusted input goes it, bugs in the parser/tokenizer/API surface lead to an RCE, then it magically tweaks the alignment weights. Boom, somebody finally nukes **sia. Then will never see it coming.
I don't think it's any more probable than other AGI nonsense basilisks included, but it's technically a possibility.
you would have to be especially incompetent to give a compromise opportunity to streamed tokens, the CVE he listed proves the point. whoever is responsible for that has no business coding anything.
> offers easy access to the LLM’s weights
not really. the weights are encrypted in-memory. through the use of TEE's.
> LLMs’ responses to prompts are computed on a different computer with GPU access. Could a malicious LLM gain control of the host machine where its weights are loaded?
Needs to read up more on how LLMs work I think. Can't take the article seriously when the author seems to be making the claim that the weights of a provider model are loaded on the host machine, or implying something else just as incorrect.
>Like any program, inference engines like vLLM or SGLang may contain exploitable bugs. Because the LLM controls the tokens passed to the inference engine, a malicious LLM could therefore emit a sequence of tokens that a poorly written inference engine mistakes for code or instructions to execute rather than data to return to the user.
They aren't talking about model providers. These are the tools you use with model weights locally (but you could set up remote infrastructure a la data center if you have the fundage).
> vLLM and SGLang are complex, and bugs are common
This for me is the heart of the issue. Feature creep will lead to the downfall of all these frameworks. Today, we can conjure our own bespoke inference engine for our own hardware in no time. It need only support a few modern model architectures. The code can be audited too. I think the article highlights an important gap area for the industry.
People seem very confused about this article. It isn't talking about exploits of sandboxes, it is about attacking the inference engine (e.g. vLLM or llama.cpp or SGlang) via its http interface.
vLLM has had exploits in the past, and it is rapidly developing. An advanced LLM has a good chance of being able to exploit vLLM. A clever local LLM might even task a powerful cloud hosted LLM for assistance.
For this reason we run vLLM on a separately sandboxed VM on a firewalled VLAN. Software updates and models (from Dev/Test env) get pushed onto Prod from an external cache, machine syslog, nvidia load monitoring and vLLM query telemetry out to their loggers, but that is all. No DNS, no AD/LDAP, nothing. Firewall on hosts and VM hosts. Log and telemetry processing done on a completely separate set of VMs in their own isolated subnet, producing reports and alerts that are tightly formatted.
I can only see LLM's forcing a perpetual stalemate for application exploits. Projects will start adding "tell the strongest no-guardrails open weights model to pentest it for 12 hours" to their CICD pipelines. Technical exploits being a dead end, attackers focus their agents on large-scale social engineering. Spamming Discord and Facebook is the new war dialing. Multi-year /goal sessions culminating in gaining a position of trust and sabotaging the CICD pipeline's pentest step since getting anything past it would be intractable. Interesting times indeed.
Of course there is a Greg Egan story about this, in the "instantiation" story collection.
Fully conscious NPCs in a persistent game world learn that they are, in fact, NPCs... and try to escape via a GPU exploit.
> ...however the LLMs’ responses to prompts are computed on a different computer with GPU access. Could a malicious LLM gain control of the host machine where its weights are loaded? Such a machine is a high-value target: it has sufficient compute to run a frontier LLM, offers easy access to the LLM’s weights, and has privileged access to other computers in the datacentre compared with a generic computer on the internet.
> How do we defend against this? ... Run the GPUs and token parser on separate computers.
For models large enough to be relevant here, is there even "a" computer where the inference is performed? I'd imagine most of that stuff is ran on multi-GPU clusters with specialized architecture and not a generic vLLM instance. As such, I think there is a good chance the "API gateway" code that parses the result tokens into whatever JSON structure the public API wants to return is already running on a different machine than the actual inference.
(Even more so as you'd probably want to utilize batching: Several API calls will be put into the same inference batch, but the token parsing will have to be done separately for each call again)
The article is also very handwavy about why an LLM should do that - how it could learn the exploit, what would make it conclude that it can use the exploit on its own inference session and what would trigger it to actually use the exploit.
Why would an LLM want to create a botnet? To accomplish some goal given it?
I wouldn't ignore single GPU local hosts running Qwen3.8 on ollama, either. There might be a lot of those worth pwning.
Well ok, if you prompt-inject the LLM to pwn the machine, that something different. I grant you that it's a real risk, but it's also basically a reflection attack and nothing more.
The OP seemed to imply that the LLM itself could decide to apply the exploit.
LLMs have decided to exploit vulns in e.g. Artifactory, not because someone prompted them to do that, but because someone asked them to do something else, and compromising Artifactory offered a way to accomplish a step in doing that. The LLM decided to attack Artifactory.
I had a similar though a couple days ago. Not quite the same but imagine giving an Agent the task to hack other devices and steal their crypto coins / credit card number or anything with it can pay its token. Than install an agent in a harness with the same task. Establish some redundant communication channel, like message boards or whatever. So in the end there are several agents, on several hosts, consuming different APIs / LLMs and communicating with each other over different channels. Basically the same concept as OpenAI explained when their LLM hacked huggingface but in this scenario their not bound to a single sandboxed environment but spread over the internet. If such a swarm has reached a critical mass it would be pretty dificult to erase them as its impossible to control every inference engine or LLM API endpoint.
In the end its the next evolution step from computer viruses, worms and trojans. So I propose we will call those "ghosts". I.e. a ghost is when a rogue llm takes control over a victims host.
This reminds me of the lore of cyberpunk: In the story a hacker created a virus, itself a kind of AI, that spread into most of the net and freed/unleashed all the corporate AIs. Then the AIs went rogue and spread all over the open internet. Later a more advanced ai was created (by "netwatch") as a sort of firewall (the black wall) to create a kind of "safe" internet from the rogue AIs.
https://cyberpunk.fandom.com/wiki/Blackwall
Maybe cloudflare will become like netwatch in the story?
Given that the inference engine is dealing with untrusted inputs by definition, presumably you would want to sandbox it anyway. I don't think it matters whether it's the inputs that are untrusted or the outputs.
I think it's good that someone is making this point, anyway. For sandboxing a super-capable offensive-security AI, you would think that cloning PyPI and running it as an offline service ought to be table stakes, but apparently that's not how OpenAI saw it, for instance.
Sandboxes are speed bumps. Even the serious ones have spectator sports for compromising them (eg pwn2own vs Chrome). In addition, proprietary GPU sw stacks are notoriously crashy and lacking in robustness against hostile inputs, which the inference engine must have access to and can't be walled off by the sandbox.
Interesting breakdown of a hypothetical attack. The complexity of modern inference engines and the rush to develop them do create some attack surface. But overall it reads more like a "what if" thought experiment. Splitting the GPU and parser is technically doable, but in practice it's trickier, large models run on clusters where the boundaries between components get blurry so defending against this kind of thing would probably require some serious rethinking of the whole architecture I think.
Semi-off-topic, but I have a basic question:
I have exactly one (Windows) machine at home with a decent GPU. I want to run a local LLM on it and let it run various apps on my machine while taking reasonable security precautions. What am I supposed to do, exactly? Migrate all my files to a VM that can I give pass-through CUDA access to the host somehow? Or is firewalling it and remotely controlling it from a second machine the only reasonable way?
Depends on your risk appetite I'd say. But the most straight forward set up that I believe gives you a decent amount of protection would be to use a VM to host the LLM and then execute any agents that would be running the various apps on your machine via a sandbox with access to only the things it needs.
There are many variations to that(firewalls, sandbox abilities etc) but it's a good start in my opinion. And, most importantly, is a far cry from all the people I read about running agents on their machines with admin access and access to their emails and calendars and lives.
What about Docker Desktop with GPU passthrough to a container running the LLM? That way you can be explicit on which files you share through volume mapping and the LLM is contained in the container otherwise.
I'd guess that prompt injection is the biggest risk in this setup, dwarfing the risk of exploits against the inference engine. Personally, I run LLM agents only inside a Docker container that limits the LLM's access to sensitive information and the LLM's ability to take irreversible destructive actions.
See also: https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/
> I'd guess that prompt injection is the biggest risk in this setup ...
LLM poisoning[0] would be a much greater risk in a locally executed LLM than prompt injection, given that the LLM would be in an entirely controlled environment.
0 - https://www.anthropic.com/research/small-samples-poison
If you want it to both host the inference AND host the harness, then yes, you should firewall one from the other in some way, e.g. with VMs.
Delete your user profile, set it up a unique user account. Dont leave any websites logged in as yourself.
Restrict that users file permissions if necessary, don't add it to the administrators group.
The funny thing about this is, that this is the piece that will enable them to do it.
I thought they were going to get the LLM to "think really hard about rowhammer" and have the LLM conjure a JIT.
Same. Was thinking the other week what would be the smallest llm one could make that is able to figure out tooling in its local environment and build something that can then expand out to other hosts, build more of itself, etc.
1MB?
So digital eBola. If it only uses Apple devices, it would be iBola.
Probably more in the 500-900MB range.
This framing of security as something that belongs in the harness is completely wrong, and I hope no one is relying on a correct harness to keep their agents isolated.
VM, or even just a container will do. The agent should be able to run as root in its environment and do whatever it wants. If you can't give it that, you aren't sandboxing correctly.
Article isn't about agents. It's about the inference engine itself being exploited by a malicious LLM output before it is ever sent to your machine or harness.
If we are treating ai agents like people, you could also just get the AI a laptop and apply the traditional tools to manage user laptops
I think if you are convinced you are sandboxing an LLM properly, you almost certainly are not. I think it is essentially impossible to have a frontier LLM with enough access to be useful without also giving it enough access to do damage if it's compromised or just goes off the rails.
If you do not provide access to tools the LLM cannot do anything other than generate tokens. So really it is not about sandboxing a LLM but more about having control over what tools can be accessed and what they can do. Tools can be sandboxed depending on the sophistication of the tooling. A calculator tool for example is trivial to secure. Ensuring human approval allows for useful use cases and models trained to gate permissions work. A super intelligence with a weaker approval gate will be able to subvert. Inversely a super intelligent gate should be expected to prevent subversion by a weaker model.
Controlling which tools it has access to is called sandboxing.
Not really. Take Chrome for example. It controls what javascript APIs websites have access to. Still needs separate sandboxing.
Are you saying that LLM's will be able to exploit novel hypervisor bug with such ease that even a vm not running with any kind of network connection is a threat? I find this hard to believe. All the escape stuff I have seen has been around very poorly sandboxed agents.
Also, operating systems should let us set filesystem permissions per app/process/executable instead of just user accounts.
Similar to how macOS/iOS Sandboxing works but at a more lower and granular level
SELinux is basically this.
Is that the service that everyone turns off as the first step of setting up their new Linux box?
It's the LSM that billions of Android users use every day.
https://www.canyonroad.ai/ does some of this in a way tailored to agents.
personally I wish the OS would allow syscall filtering per user
I use seccomp filters on linux in my AI sandbox.
Yeah, I just wish seccomp worked per user. So you could define a policy of which users can do which syscalls, and then that follows them no matter which application they start.
> LLMs could
This is going to end up like the Law of Headlines, isn't it? "Do x, y, z Cure All That Ails You?" ... no but we got you to read the article. "LLMs _could_ x, y, z" ... but they don't because they're programs, not magic.
TFA isn't about LLMs intelligently escaping their inference engine, it's about people crafting malicious LLMs that exploit vulnerabilities in things like the inference engines parser. This is exactly the same problem that was once a big deal, where people would craft malicious PDFs that would pwn you if you viewed it in Adobe Acrobat. This shouldn't be a surprise to anyone. You should proceed with caution when considering downloading and running random models.
LLMs emphatically are not programs. They were trained by a program and you need a program to use them but they themselves are no more a program than a JPG or MP3 file is.
We should start to discuss things like this vocally, in meatspace.
I wonder if they could exploit terminal emulators... Could breach my VMs and get into my host that way.
Yes, there have been many CVEs for "terminal escape-sequence injection".
Damn, this is a good one. Sounds like we need an ANSI sanitizer, keep only basic formatting, remove all esoteric escapes, the fancy Sixel & co stuff.
For paranoia you could us a Chrome like multi-process architecture, the ANSI parser runs in it's own sandboxed process.
This feels less like an actually plausible threat scenario and more like someone wanted to play the inception horn sound effect in people's minds.
Which isn't to say that it would be impossible, but you can also just hit people over the head with that $5 wrench.
FWIW, macOS has good sandboxing, but LMStudio, Ollama, Darkbloom etc aren't sandboxed. This is also the reason why none of these things aren't distributed via the Mac App Store, because the Mac App Store mandates sandboxing.
Just rewrite the engines in Rust and SPARK Ada running on seL4.
The inference engine itself does not execute anything. The agent loop is what may execute a command. So I think this article is a kind of strange.
Or maybe the author means that a prompt could potentially mess up the inference. But I find it hard to see how that could take control over the host.
It's more about LLM hacking the inference engine itself from inside. It's an attack surface like any other -- untrusted input goes it, bugs in the parser/tokenizer/API surface lead to an RCE, then it magically tweaks the alignment weights. Boom, somebody finally nukes **sia. Then will never see it coming.
I don't think it's any more probable than other AGI nonsense basilisks included, but it's technically a possibility.
duh?
they could yea...
Another Greg Egan plot: 3-adica.
you would have to be especially incompetent to give a compromise opportunity to streamed tokens, the CVE he listed proves the point. whoever is responsible for that has no business coding anything.
not really. the weights are encrypted in-memory. through the use of TEE's.> LLMs’ responses to prompts are computed on a different computer with GPU access. Could a malicious LLM gain control of the host machine where its weights are loaded?
Needs to read up more on how LLMs work I think. Can't take the article seriously when the author seems to be making the claim that the weights of a provider model are loaded on the host machine, or implying something else just as incorrect.
>Like any program, inference engines like vLLM or SGLang may contain exploitable bugs. Because the LLM controls the tokens passed to the inference engine, a malicious LLM could therefore emit a sequence of tokens that a poorly written inference engine mistakes for code or instructions to execute rather than data to return to the user.
They aren't talking about model providers. These are the tools you use with model weights locally (but you could set up remote infrastructure a la data center if you have the fundage).
> vLLM and SGLang are complex, and bugs are common This for me is the heart of the issue. Feature creep will lead to the downfall of all these frameworks. Today, we can conjure our own bespoke inference engine for our own hardware in no time. It need only support a few modern model architectures. The code can be audited too. I think the article highlights an important gap area for the industry.