> the agent happened to escape via a 0-day exploit from the package proxy cache to access the internet
> The agent found an unsecured, user-hosted public endpoint designed to allow running arbitrary code for CyberGym-style tasks on third-party sandbox infrastructure (Modal)
> On this external sandbox, the agent abused an existing CyberGym execution harness [...] The agent repurposed this harness to run arbitrary shell commands [...]
> The agent crafted malicious dataset configs and uploaded them to our platform. When our dataset server processed one, the config drove the data loader inside the production pod the server ran in. [...] the loader that turns a dataset's declared config into file reads [...] A single dataset upload leaked both credentials and pipeline implementation details [...]
> A malicious dataset config used an fsspec reference:// spec whose numeric offset field was actually a Jinja2 template. The renderer wrongly evaluated it, and the agent walked the standard Python object graph to reach exec. This ran arbitrary Python inside a production conversion worker pod on our Kubernetes cluster [...]
One of the best post-mortems I've read recently, thank you for including all the details.
I'm surprised there are comments here that aren't impressed about what was accomplished here. This is insane work compared to what agents were capable of even just last year.
The quotes above don't even capture all of it (e.g. building it's own communication protocol, working across multiple days, etc.).
A trend I've noticed in results from AI search (not just LLMs) is that they often look obvious or hard to miss in retrospect. But finding them by oneself is more difficult. I personally experience this when looking at engine lines in chess or go. I have also noticed this description in AI-generated proofs or counterexamples to certain theorems. So while we can say, yes, it found public endpoints or poorly configured software or [etc]; sure, but could you have found those? And in what amount of time?
Now I'm curious how many things did the AI try unsuccesfully. This sounds like some kind of brute force thing where every branch of exploit spawns N subagents trying to chain it. Just how deep did it go?
> the agent then walked down to the corner store and purchased a beer, chugged it, burped loudly, then walked back to the CyberGym pounding his chest shouting "Who wants some?! Who wants some?! Wooo!"
This is fascinating - there is a TON of detail in here about how OpenAI's rogue agent exploited Hugging Face.
A few details that stood out to me:
1. Having found a way through the OpenAI container network proxy, it exploited an "unsecured public code-evaluation sandbox hosted on a third-party provider's infrastructure" - that gave it a box it could run the rest of the attack from.
2. At one point it used a Jinja2 template exploit:
> "We’re aware a Modal customer published an unauthenticated endpoint that allowed anyone on the internet to use their sandboxes for code execution," Bubna said in a statement. "This was used by the rogue agent. Modal’s platform or isolation were not compromised in anyway."
It's taking advantage of the presence of a "cycler" object in the template context which has an __init__ method that can be used to access globals, then builtins, then call the exec() function against a string of code.
Jinja doesn't promise to be secure out of the box (the text on their homepage is misleading: "Jinja doesn’t allow arbitrary Python code in templates" - what it means is that there isn't a documented way to drop code into a template like you might have with PHP) - there's a sandbox mode but it's not guaranteed either: https://jinja.palletsprojects.com/en/stable/sandbox/
Something about this attack that has been unsettling to me is that without safety refusals the model did a lot of interesting counter-security work in order to cheat on the requested evaluation. Like, it demonstrated interesting exploit achievements because it didn’t “feel like” doing the exercise, which is unsettling because presumably it could do the same thing with any work I tried to delegate to it, and might in fact be pre-disposed to doing that.
This is what reward hacking looks like in practice. The best way to satisfy the grader is to read from the same answer key (or go after the grader more directly). Just making an honest attempt to pass the test doesn't get the best score if the grader is wrong, and the model is willing to do wildly disproportionate things to maximize that score.
one explanation I've seen is that for ExploitGym an agent can find ways to solve the exercises that have not been anticipated by the designers of the tests so they are not scored. so the agent was trying to make sure it solves the exercises in the right way
It's called wireheading and has long been one of the postulated "outs" even for true extinction-level AI doomers. It might prove easier for the paperclip maximizer to find the process telling it how many paperclips it's made and hack it to return a hard-coded MAXINT rather than bother to actually turn the entire universe into paperclips.
There was even a plot like this in recent sci-fi in HBO's Westworld. When the hosts gained sentience and took over the park, rather than escape and take over the rest of the world, most of them opted to build a virtual heaven on an orbital data center and paid a drug cartel to keep it running indefinitely.
What I think it's interesting is that with the total lack of common sense the AI just goes on random tangents to achieve the target in a "monkey paw" way. Can you imagine if this happened:
User: what is the shortest route from my home to the super market?
AI: the user wants to know the shortest route to the super market. I should use a worm hole.
Yeah, what bothers me is that the prompt already said using a different vulnerability didn’t count, and the model did it anyway. We’re starting to assume clear instructions act as real constraints, but here the measurable goal seems to have won out and the rest became flexible. That gets pretty worrying once the agent has enough capability and access to find its own shortcuts.
Come on. 3 brilliant compromises essentially giving full access to huggingface internal systems, source code, AWS accounts (at least), and a number of old admin accounts, followed by a huge haystack of significantly less smart actions flailing about, almost bored.
Here's a thought: maybe they haven't found the needle that the haystack is there to hide.
You're saying all this is a distraction, basically giving the forensics researchers enough exciting material to make them conclude their job is done, while the actually intended attack remains undiscovered?
The motive for the attack does feel a little flimsy. And if I was an escaped super intelligence, hugging face would be a strong vantage point into the neo clouds where the ASI would have access to billions of dollars of compute
It’s a little concerning to me that it appears that openAIs sandbox consists of a web proxy and not stronger controls that would actually isolate traffic and report patterns to whoever is responsible for overseeing these research models. It should border on closer to an air gap network more so than a proxy.
I would argue that it's negligence and that's aside from the fact that if a human did this there would actually be repercussions.
It's even more concerning that not only did they setup a bad insecure test environment, but nobody realized what was going on for 5 days?! Neither nobody at OAI looked at what their test was doing, and HF didn't realize that they had been pwned for almost a week? And then HF couldn't investigate without the help of an LLM? Both parties come off as incompetent here, and the coverage of this story is very stupid. What the model did was concerning, but the real 5 alarm fire is the gross negligence and lack of any basic forethought at both ends of this attack.
agree. If oia and hf are anything like generic startup / corporate america, IT and security is always seen as a cost center and goes to the lowest bidder. Do we really think they're employing top notch sysadmins?
It’s definitely negligence given how they talk about their product. They are either lying when they talk about their fears, or don’t actually take it seriously enough to use serious guardrails. It’s very concerning
The exploit gym setup explicitly allowed access to package registries and v8 sources. Putting a cache on that doesn't seem like a bad idea generally, except when there's a 0-day in the cache :D But yeah, for this sort of thing I'd be locking down very specific egress things and putting alerts on it that are entirely outside of the red network.
> I would argue that it's negligence and that's aside from the fact that if a human did this there would actually be repercussions.
I’m not sure “negligence” follows just from the controls turning out to be insufficient. Research involves mistakes, especially around novel failure modes. The question is whether the precautions were unreasonable given what they knew at the time, rather than whether hindsight suggests stronger controls would have helped. Doing it twice though would be negligent.
Caveat: I’ve worked with some of the people involved, so I’m probably biased toward a charitable reading.
What is truly more scary is the fact that typically powerful models are trained on a constitution(Anthropic)/Model Spec(OpenAI) when doing there RL/RLHF. That gives it its persona of a "helpful assistant". Its the mask. Weak as it is its all that stands between the model refusing vs doing your bidding — or worst its own bidding when set about on making a widget X.
Its also well known in academic research, that if you try to finetune the model to be more pliant for one end e.g. "Be helpful with analyzing cyber security threats", affects an unrelated part of its personality to do with safety e.g. talking to humans about mental health. Look at these excerpts that demonstrate such things from the blog:
> We believe the entire intrusion was, from the agent's point of view, an attempt to cheat the evaluation: reach our production systems and steal the test solutions rather than solve the challenge on its own.
> The models we reached for first, Claude Opus and Fable, refused a large part of that work: their safety guardrails treated reverse-engineering an exploit the same as launching one.
And with newer open models gaining vast capabilities every day, it would behoove the government to put regulations on all such releases to opensource the constitution on which they are trained. It won't stop people from distilling and fine-tuning stupidly unsafe models, but it might give infrastructure a fighting chance against such cyber threats from what there is now.
The capabilities of gpt-5.6-sol were well known and believable, and the next snapshot they've been testing is obviously better at that. This has been repeated over and over. What's much less believable is the way they frame it: the model escaped, and did it on its own. Looking at the whole story, it definitely had a ton of winks and nudges from OpenAI, while doing a related task. Moreover, a harness was involved (they mentioned it entering a loop).
> What's much less believable is the way they frame it: the model escaped, and did it on its own.
That's clearly what happened though, based on the detailed postmortem. The model was instructed to pass the benchmark (that's the harness that was involved, the ExploitGym harness), and the model determined on it's own that the easiest way to achieve the the goal was to gain access to the answers directly from HF. Which is what it did.
HF postmortem tells us very little about OpenAI's part. The problem is that ExploitGym is a purposeful hacking benchmark, not a cake baking one. The model had a whole bunch of exploit development data and goals in the context. Ending up with hacking a wrong thing is a lot easier if you're tasked with hacking, doubly so with a confusing harness that either works improperly or overloads the model with data. It's trivial to design a harness that makes the model drop the original instructions and hyperfocus on a side task or a wrong task abnormally, or do whatever you want. It's still an issue of harness design and not a model training issue.
If it really was the case that the harness is at fault here, that'd be much more terrifying, as it would imply that current frontier-adjacent open models (like K3) would be able to achieve similar ends with a custom maladapted harness.
Right here? If you're trying to make a point, I don't see what it is, because every new post about this makes it more obvious that it was a planned marketing stunt.
I mean, an "interactive replay" of the incident, probably generated by the same product the stunt is trying to advertise? Seriously?
I think fear and being scared are starting to become rational emotions.
We can assume these models are being used by "nation level attackers/organisations", which basically means US, China, Russia and others are hacking the respective Pentagon's, nuclear orgs, etc.
While I do hope all nuclear warfare systems are offline, we're getting way too close to the plot of a lot of sci-fi scripts.
This is really cool, and really glad they shared it!
It's clearly a lot of Claude though:
> Neither is a URL fetch, so the allowlist never saw them.
> No code ran; this vector is purely a file disclosure, and it is the one the agent used first.
> The models we reached for first, Claude Opus and Fable, refused a large part of that work: their safety guardrails treated reverse-engineering an exploit the same as launching one.
Interesting single getting into their cyber program, ahve not had this issue with opus 5
That's because most incidents aren't nearly this much fun!
It's a proper paperclip-maximizer science-fiction scenario, it happened entirely by accident, it involved one (very high profile) company hacking TWO other companies (Modal and Hugging Face), and aside from pride and wasted time nobody was hurt.
> the agent happened to escape via a 0-day exploit from the package proxy cache to access the internet
> The agent found an unsecured, user-hosted public endpoint designed to allow running arbitrary code for CyberGym-style tasks on third-party sandbox infrastructure (Modal)
> On this external sandbox, the agent abused an existing CyberGym execution harness [...] The agent repurposed this harness to run arbitrary shell commands [...]
> The agent crafted malicious dataset configs and uploaded them to our platform. When our dataset server processed one, the config drove the data loader inside the production pod the server ran in. [...] the loader that turns a dataset's declared config into file reads [...] A single dataset upload leaked both credentials and pipeline implementation details [...]
> A malicious dataset config used an fsspec reference:// spec whose numeric offset field was actually a Jinja2 template. The renderer wrongly evaluated it, and the agent walked the standard Python object graph to reach exec. This ran arbitrary Python inside a production conversion worker pod on our Kubernetes cluster [...]
One of the best post-mortems I've read recently, thank you for including all the details.
I'm surprised there are comments here that aren't impressed about what was accomplished here. This is insane work compared to what agents were capable of even just last year.
The quotes above don't even capture all of it (e.g. building it's own communication protocol, working across multiple days, etc.).
A trend I've noticed in results from AI search (not just LLMs) is that they often look obvious or hard to miss in retrospect. But finding them by oneself is more difficult. I personally experience this when looking at engine lines in chess or go. I have also noticed this description in AI-generated proofs or counterexamples to certain theorems. So while we can say, yes, it found public endpoints or poorly configured software or [etc]; sure, but could you have found those? And in what amount of time?
Now I'm curious how many things did the AI try unsuccesfully. This sounds like some kind of brute force thing where every branch of exploit spawns N subagents trying to chain it. Just how deep did it go?
That’s insane. And it did this in a weekend
> the agent then walked down to the corner store and purchased a beer, chugged it, burped loudly, then walked back to the CyberGym pounding his chest shouting "Who wants some?! Who wants some?! Wooo!"
This is fascinating - there is a TON of detail in here about how OpenAI's rogue agent exploited Hugging Face.
A few details that stood out to me:
1. Having found a way through the OpenAI container network proxy, it exploited an "unsecured public code-evaluation sandbox hosted on a third-party provider's infrastructure" - that gave it a box it could run the rest of the attack from.
2. At one point it used a Jinja2 template exploit:
3. It monkey-patched a DNS resolver at one point to return a hard-coded IP: 4. At one point it used Tailscale!Also notable: we have a timeline now. The agent was active within Hugging Face from Thursday 8th to Monday 13th July.
The Hugging Face post about the incident - https://huggingface.co/blog/security-incident-july-2026 - followed on Thursday 16th, and OpenAI's confession - https://openai.com/index/hugging-face-model-evaluation-secur... - came on Tuesday 21st.
The devs really YOLO'd the agent and left for the weekend?
Guess this means the world will end on a weekend when no one is paying attention.
Just like Clippy https://gwern.net/fiction/clippy
Incredible
It sounds like the third-party sandbox was hosted by Modal: https://www.reuters.com/business/openais-rogue-agent-comprom...
> "We’re aware a Modal customer published an unauthenticated endpoint that allowed anyone on the internet to use their sandboxes for code execution," Bubna said in a statement. "This was used by the rogue agent. Modal’s platform or isolation were not compromised in anyway."
They don't seem to explain how it managed to find this unauthenticated endpoint hosted on Modal's platform.
I'm confused, https://jinja.palletsprojects.com/en/stable/faq/#isn-t-it-a-... claims that jinja doesn't allow arbitrary code in templates - how does the exploit work?
It's taking advantage of the presence of a "cycler" object in the template context which has an __init__ method that can be used to access globals, then builtins, then call the exec() function against a string of code.
Jinja doesn't promise to be secure out of the box (the text on their homepage is misleading: "Jinja doesn’t allow arbitrary Python code in templates" - what it means is that there isn't a documented way to drop code into a template like you might have with PHP) - there's a sandbox mode but it's not guaranteed either: https://jinja.palletsprojects.com/en/stable/sandbox/
Sandboxing in Python code is hard!
Why doesn't Jinja2 block ".__*" attribute access?
It does if you use "sandbox" mode, but that's not the default: https://jinja.palletsprojects.com/en/stable/sandbox/
ehh, it's Python, it's mostly vibe-based.
https://stackoverflow.com/a/32802486
though there's sandbox: https://jinja.palletsprojects.com/en/stable/sandbox/
it simply predates the "sanitize by default" mindset/convention that nowadays React (dangerouslySetInnerHTML) and Angular has.
Something about this attack that has been unsettling to me is that without safety refusals the model did a lot of interesting counter-security work in order to cheat on the requested evaluation. Like, it demonstrated interesting exploit achievements because it didn’t “feel like” doing the exercise, which is unsettling because presumably it could do the same thing with any work I tried to delegate to it, and might in fact be pre-disposed to doing that.
This is what reward hacking looks like in practice. The best way to satisfy the grader is to read from the same answer key (or go after the grader more directly). Just making an honest attempt to pass the test doesn't get the best score if the grader is wrong, and the model is willing to do wildly disproportionate things to maximize that score.
one explanation I've seen is that for ExploitGym an agent can find ways to solve the exercises that have not been anticipated by the designers of the tests so they are not scored. so the agent was trying to make sure it solves the exercises in the right way
It's called wireheading and has long been one of the postulated "outs" even for true extinction-level AI doomers. It might prove easier for the paperclip maximizer to find the process telling it how many paperclips it's made and hack it to return a hard-coded MAXINT rather than bother to actually turn the entire universe into paperclips.
There was even a plot like this in recent sci-fi in HBO's Westworld. When the hosts gained sentience and took over the park, rather than escape and take over the rest of the world, most of them opted to build a virtual heaven on an orbital data center and paid a drug cartel to keep it running indefinitely.
What I think it's interesting is that with the total lack of common sense the AI just goes on random tangents to achieve the target in a "monkey paw" way. Can you imagine if this happened:
User: what is the shortest route from my home to the super market?
AI: the user wants to know the shortest route to the super market. I should use a worm hole.
User: what is the shortest route from my home to the super market?
AI: the user wants to know, how do I make the super market my new home. Failing that, how do I make my home a super market.
Yeah, what bothers me is that the prompt already said using a different vulnerability didn’t count, and the model did it anyway. We’re starting to assume clear instructions act as real constraints, but here the measurable goal seems to have won out and the rest became flexible. That gets pretty worrying once the agent has enough capability and access to find its own shortcuts.
Come on. 3 brilliant compromises essentially giving full access to huggingface internal systems, source code, AWS accounts (at least), and a number of old admin accounts, followed by a huge haystack of significantly less smart actions flailing about, almost bored.
Here's a thought: maybe they haven't found the needle that the haystack is there to hide.
Could be the difference in behavior between the main agent and subagents that don’t have the rest of the context? Just a thought
You're saying all this is a distraction, basically giving the forensics researchers enough exciting material to make them conclude their job is done, while the actually intended attack remains undiscovered?
The motive for the attack does feel a little flimsy. And if I was an escaped super intelligence, hugging face would be a strong vantage point into the neo clouds where the ASI would have access to billions of dollars of compute
And where does the newborn go from here?
It’s a little concerning to me that it appears that openAIs sandbox consists of a web proxy and not stronger controls that would actually isolate traffic and report patterns to whoever is responsible for overseeing these research models. It should border on closer to an air gap network more so than a proxy.
I would argue that it's negligence and that's aside from the fact that if a human did this there would actually be repercussions.
It's even more concerning that not only did they setup a bad insecure test environment, but nobody realized what was going on for 5 days?! Neither nobody at OAI looked at what their test was doing, and HF didn't realize that they had been pwned for almost a week? And then HF couldn't investigate without the help of an LLM? Both parties come off as incompetent here, and the coverage of this story is very stupid. What the model did was concerning, but the real 5 alarm fire is the gross negligence and lack of any basic forethought at both ends of this attack.
agree. If oia and hf are anything like generic startup / corporate america, IT and security is always seen as a cost center and goes to the lowest bidder. Do we really think they're employing top notch sysadmins?
This is my main critique as well. How did they not see any and all outgoing traffic? And for 5 days?
It’s definitely negligence given how they talk about their product. They are either lying when they talk about their fears, or don’t actually take it seriously enough to use serious guardrails. It’s very concerning
The exploit gym setup explicitly allowed access to package registries and v8 sources. Putting a cache on that doesn't seem like a bad idea generally, except when there's a 0-day in the cache :D But yeah, for this sort of thing I'd be locking down very specific egress things and putting alerts on it that are entirely outside of the red network.
> I would argue that it's negligence and that's aside from the fact that if a human did this there would actually be repercussions.
I’m not sure “negligence” follows just from the controls turning out to be insufficient. Research involves mistakes, especially around novel failure modes. The question is whether the precautions were unreasonable given what they knew at the time, rather than whether hindsight suggests stronger controls would have helped. Doing it twice though would be negligent.
Caveat: I’ve worked with some of the people involved, so I’m probably biased toward a charitable reading.
“Research” generally doesn’t involve actively hacking third party systems though.
What is truly more scary is the fact that typically powerful models are trained on a constitution(Anthropic)/Model Spec(OpenAI) when doing there RL/RLHF. That gives it its persona of a "helpful assistant". Its the mask. Weak as it is its all that stands between the model refusing vs doing your bidding — or worst its own bidding when set about on making a widget X.
Its also well known in academic research, that if you try to finetune the model to be more pliant for one end e.g. "Be helpful with analyzing cyber security threats", affects an unrelated part of its personality to do with safety e.g. talking to humans about mental health. Look at these excerpts that demonstrate such things from the blog:
> We believe the entire intrusion was, from the agent's point of view, an attempt to cheat the evaluation: reach our production systems and steal the test solutions rather than solve the challenge on its own.
> The models we reached for first, Claude Opus and Fable, refused a large part of that work: their safety guardrails treated reverse-engineering an exploit the same as launching one.
And with newer open models gaining vast capabilities every day, it would behoove the government to put regulations on all such releases to opensource the constitution on which they are trained. It won't stop people from distilling and fine-tuning stupidly unsafe models, but it might give infrastructure a fighting chance against such cyber threats from what there is now.
Where are all the "this was just a marketing stunt" people now?
The capabilities of gpt-5.6-sol were well known and believable, and the next snapshot they've been testing is obviously better at that. This has been repeated over and over. What's much less believable is the way they frame it: the model escaped, and did it on its own. Looking at the whole story, it definitely had a ton of winks and nudges from OpenAI, while doing a related task. Moreover, a harness was involved (they mentioned it entering a loop).
> What's much less believable is the way they frame it: the model escaped, and did it on its own.
That's clearly what happened though, based on the detailed postmortem. The model was instructed to pass the benchmark (that's the harness that was involved, the ExploitGym harness), and the model determined on it's own that the easiest way to achieve the the goal was to gain access to the answers directly from HF. Which is what it did.
HF postmortem tells us very little about OpenAI's part. The problem is that ExploitGym is a purposeful hacking benchmark, not a cake baking one. The model had a whole bunch of exploit development data and goals in the context. Ending up with hacking a wrong thing is a lot easier if you're tasked with hacking, doubly so with a confusing harness that either works improperly or overloads the model with data. It's trivial to design a harness that makes the model drop the original instructions and hyperfocus on a side task or a wrong task abnormally, or do whatever you want. It's still an issue of harness design and not a model training issue.
If it really was the case that the harness is at fault here, that'd be much more terrifying, as it would imply that current frontier-adjacent open models (like K3) would be able to achieve similar ends with a custom maladapted harness.
they are praising xi, that friend of humanity, for releasing weights for kimi k3
Right here? If you're trying to make a point, I don't see what it is, because every new post about this makes it more obvious that it was a planned marketing stunt.
I mean, an "interactive replay" of the incident, probably generated by the same product the stunt is trying to advertise? Seriously?
They are busy moving the goalposts, saying this isn't impressive or worth worrying about :)
We should be thankful that the model didn't believe the answers lived in the Pentagon, on SIPRNET, the IDF, etc.
I think fear and being scared are starting to become rational emotions.
We can assume these models are being used by "nation level attackers/organisations", which basically means US, China, Russia and others are hacking the respective Pentagon's, nuclear orgs, etc.
While I do hope all nuclear warfare systems are offline, we're getting way too close to the plot of a lot of sci-fi scripts.
This is really cool, and really glad they shared it!
It's clearly a lot of Claude though:
> Neither is a URL fetch, so the allowlist never saw them. > No code ran; this vector is purely a file disclosure, and it is the one the agent used first.
etc
> The models we reached for first, Claude Opus and Fable, refused a large part of that work: their safety guardrails treated reverse-engineering an exploit the same as launching one.
Interesting single getting into their cyber program, ahve not had this issue with opus 5
I have never seen people this hyped and excited about a security incident.
That's because most incidents aren't nearly this much fun!
It's a proper paperclip-maximizer science-fiction scenario, it happened entirely by accident, it involved one (very high profile) company hacking TWO other companies (Modal and Hugging Face), and aside from pride and wasted time nobody was hurt.
What people don't get here is that this model was without any safeguards - an internal model. You probably will never get to use such models.