Rendered at 22:50:37 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
wxw 3 hours ago [-]
> 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.).
sobellian 3 hours ago [-]
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?
jvuygbbkuurx 2 hours ago [-]
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?
sourdecor 52 minutes ago [-]
I keep waiting for an AI to exfiltrate itself. That is going to be cool to read about.
tccole 3 hours ago [-]
That’s insane. And it did this in a weekend
limecherrysoda 3 hours ago [-]
> 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!"
cindyllm 3 hours ago [-]
[dead]
simonw 1 days ago [-]
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."
onesociety2022 2 hours ago [-]
They don't seem to explain how it managed to find this unauthenticated endpoint hosted on Modal's platform.
simonw 52 minutes ago [-]
Modal are a hosting provider. It sounds to me like someone building on Modal deployed their own product that had an unauthenticated endpoint that could be used to launch and interact with a container.
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/
it simply predates the "sanitize by default" mindset/convention that nowadays React (dangerouslySetInnerHTML) and Angular has.
SaucyWrong 20 hours ago [-]
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.
zmj 2 hours ago [-]
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.
CrimsonRain 21 minutes ago [-]
So best course of action for ai to get best rating after you prompt something is for it to hire a gunman to hold a gun on your head to press that like button on its reply and then shoot you anyways.
koyote 23 minutes ago [-]
This is definitely weird and I'd love to see the prompt.
Why would the model spend 4 days hacking into a machine if it is clever enough to just 'solve' the issue given?
So either the AI is actually not very clever or useful ("Write fizz-buzz" - "Sure, let me just invent a new programming language first"). or the prompt was nudging it towards such a scenario.
vuciuc 3 hours ago [-]
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
nonameiguess 2 hours ago [-]
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.
AlienRobot 2 hours ago [-]
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.
genericone 2 hours ago [-]
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.
sobellian 53 minutes ago [-]
Just don't ask what the shortest path is to food.
joshka 15 hours ago [-]
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.
MagicMoonlight 1 hours ago [-]
[dead]
spwa4 13 hours ago [-]
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.
dgellow 5 hours ago [-]
Could be the difference in behavior between the main agent and subagents that don’t have the rest of the context? Just a thought
TeMPOraL 3 hours ago [-]
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?
vavos 2 hours ago [-]
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
puchatek 2 hours ago [-]
And where does the newborn go from here?
llama052 23 hours ago [-]
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.
strange_quark 3 hours ago [-]
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.
physPop 2 hours ago [-]
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?
baq 1 hours ago [-]
Repercussions is a polite way of putting it.
…as is your ‘a little concerning’ comment about the eval infra. It probably didn’t occur to them that they have something capable of doing what it did - despite Anthropic writing a very similar story in the Fable model card, except it wasn’t a felony there. Maybe.
One does have to wonder what else OpenAI didn’t catch.
dgellow 5 hours ago [-]
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
joshka 15 hours ago [-]
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.
llama052 7 hours ago [-]
“Research” generally doesn’t involve actively hacking third party systems though.
Ancalagon 2 hours ago [-]
This is my main critique as well. How did they not see any and all outgoing traffic? And for 5 days?
3 hours ago [-]
simoneree 15 hours ago [-]
[dead]
ghm2199 7 hours ago [-]
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.
ghm2199 7 hours ago [-]
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.
6 hours ago [-]
cmiles8 6 hours ago [-]
It’s an interesting post mortem and situation but it seems to say more about the weakness of the Hugging Face architecture than it does about the strength of the models.
This mostly reads like script kiddie style hacking, not some state actor black-ops stuff.
cobolcomesback 4 hours ago [-]
The overwhelming majority of web app security issues, even the really complicated and sexy sounding ones, at their core are just “script kiddie style hacking”. Injection attacks and RCEs like this Jinja injection may seem “script kiddie” but they are a dime-a-dozen.
It is folly to assume that most companies don’t have similar weaknesses somewhere in their stacks. In the past, they were just sufficiently buried that they would get noticed and fixed before an attacker found them. With AI able to probe at thousands of these potential weaknesses per second… that’s not so true anymore.
spiresofagartha 4 hours ago [-]
This hasn't been my experience both searching and reading vulnerabilities. Hackerone is public for some companies and vulnerabilities and you can see how contrived exploits can get. Even before AI, Burp Suite Pro was enough to get the so-called script kiddie style hacking.
nickpsecurity 4 hours ago [-]
The commenter may nit be assuming that. Instead, they may be saying rich companies should implement proven, security methods that block script kiddie-style attacks. HughingFace apparently isn't doing that.
That's how I read the comment.
cmiles8 3 hours ago [-]
Correct. Basic security practices would have stopped most of this. Thus what is most insightful here is less that the attack was sophisticated but more that Hugging Face was running a setup that, by standards that are the norm elsewhere, would appear amateurish.
pyth0 3 hours ago [-]
You might be able to say that HF made some amateurish mistakes, but this attack certainly was fairly sophisticated. It relied a number of separate exploits or vulnerabilities that it combined and used over the course of multiple days. Calling it otherwise seems to be downplaying it.
4 hours ago [-]
0xDEAFBEAD 6 hours ago [-]
Didn't the model initially obtain internet access by discovering a zero-day vulnerability?
In any case, I would guess that a lot of unicorn startups like HuggingFace could be hacked by a sufficiently determined script kiddie working at 100x speed. The practical implications of a coming AI hacking wave could be large, even if agents are just doing grunt work really fast. Most organizations suck at security.
Seems to me that the most likely scenario is: Black hats are currently tuning the recent Kimi release for this type of work, and we'll see a flood of similar attacks within the next few months. (Why would this not happen?) Note that regulation is useless here, because black hats don't give a crap about regulators!
mr_mitm 4 hours ago [-]
It's a zero day in some caching proxy who may not have had that threat model in mind. I guess its primary purpose is caching packages, not restricting internet access.
That's a far cry from finding a zero day in a hypervisor or even Docker, which OpenAI conveniently left out in their first statement by simply calling it a "sandbox".
0xDEAFBEAD 4 hours ago [-]
I can't say for sure, but intuitively this type of vulnerability seems like it could be handy for an attacker trying to move through a target network? In any case, JFrog seems to be taking the situation quite seriously
That is the threat model, that is supposed to be safe isolated access in a vpc that isolated applications can access but it has external access.
cmiles8 3 hours ago [-]
The deeper question though is why is someone running a model that’s supposedly so dangerous in an environment that can even get access to the internet. That just appears reckless.
There are ways to test supposedly dangerous things. What was on display here looks more amateur hour than serious testing.
nickpsecurity 4 hours ago [-]
I wonder if whatever had the zero day was written in a memory-safe language with strong authentication and a secure parser. Such were the recommendations to stop many 0-days before GPT-2 was invented.
If it had poor security, the attack would be both evidence of poor security and proof that the agents can compromise poor security which might still be amazing.
justincormack 3 hours ago [-]
Its likely written in Java, so memory safe.
PunchyHamster 5 hours ago [-]
In OpenAI sandbox. which was probably vibe coded. That is to say it is probably far easier to achieve than escaping something more battle tested.
> Note that regulation is useless here, because black hats don't give a crap about regulators!
I'd argue more than useless and actively harmful, as you get denied access to tools that black hats use against you.
0xDEAFBEAD 5 hours ago [-]
"To gain Internet access, the models identified and exploited a previously unknown zero-day vulnerability in Artifactory, a package registry cache proxy. We disclosed this vulnerability, along with other Artifactory vulnerabilities our models identified as part of our review, to the vendor."
Based on some quick Googling, Artifactory was created by a software supply chain company called JFrog, which is publicly traded and valued at around $9 billion.
>I'd argue more than useless and actively harmful, as you get denied access to tools that black hats use against you.
Indeed. Still, perhaps the US and China can coordinate to stop releasing new models with even more advanced capabilities. Such coordination could be bad for big AI companies in the US, since they've been making money by offering more powerful models than the open-weight stuff which comes out of China. Any halt to the AI race will create equalization and erode their competitive edge. Speaking for myself, I personally wouldn't mind seeing US AI companies take a financial hit from such an an international agreement.
plandis 3 hours ago [-]
I wonder how current models would handle something like the xz backdoor? Could they perform the social engineering and technical aspects to make something like that happen?
TeMPOraL 3 hours ago [-]
The elephant in the room is still that this was done by an LLM and not by a human. Script kiddie or not, in this reality, in the year of our lord 2026, people still call SOTA LLMs "useless", "scams", and "stochastic parrots".
AlienRobot 2 hours ago [-]
It's only script kiddie hacking if you get the exploit from someone else. If you program it yourself, it's normal hacking.
maxnevermind 52 minutes ago [-]
> the agent happened to escape via a 0-day exploit from the package proxy cache to access the internet
I have same familiarity with JFrog Artifactory(which they used in this case) and I'm confused with how you can gain access to the internet through it, even in theory. Isn't JFrog just exposed with a basic REST API for package upload/extraction? I'm not a security expert but how is that possible to get internet access through simple GET/PUT request API?
wonnage 44 minutes ago [-]
The package cache proxy is usually used to fetch from the public repository (npm, rubygems, etc.) so I think it could be feasible to craft some package metadata to trick it into GETing unexpected things. PUT/POST could be possible via attempting to publish
firasd 6 hours ago [-]
Bottom line seems to be: So Huggingface has a feature where you can upload a dataset. The agent used that feature to insert an arbitrary payload into their environment. Now their environment also supports templates inside the dataset and the template ended up being evaluated into executable code
estearum 3 hours ago [-]
I don't think the "bottom line" of "a model escaped containment and then successfully attacked someone else's production infrastructure" is a description of the specific vulnerability and exploit.
That's like saying the key takeaway of 9/11 was that skyscrapers can be severely damaged by airplane collisions.
kfarr 6 hours ago [-]
If, like me, you're more interested in the actual exploit instead of the meta narrative, here's a relevant excerpt for the initial intrusion:
> Stage 1: reaching a launchpad by chaining through other parties' infrastructure. As documented in OpenAI's blog post (https://openai.com/index/hugging-face-model-evaluation-secur...), during an internal capability evaluation on OpenAI's platform, the agent escaped its sandbox by exploiting a zero-day in the package registry cache proxy, one of its primary permitted network egress with internet, then abused a public code-evaluation harness hosted by a user of a third-party infrastructure provider. It was able to run commands as root/admin on that external sandbox and used it as a control, staging, and egress base for the entire campaign: it acted as an external launchpad for the agent. That compromised external sandbox had no direct network path to our cluster; however, during our counter-intervention to stop the intrusion, we were able to retrieve some of the agent logs from this external launchpad. We used those to reconstruct the present forensic analysis.
NitpickLawyer 6 hours ago [-]
The initial escape is not that interesting, IMO. It's an exploit in a testing sandbox, sure, but it's expected to happen (or at least it should have been expected, that's why you airgap pentesting sandboxes). That is, the model is expected to try and find ways around limitations in its running environment. You kinda want that.
The fact that at some point it "decided" to focus on external resources (i.e. hf) and successfully penetrated everything in its path from sandbox to hf's infra, all the way through internal VPNs, to access internal code repos, that's the interesting part. Especially because it chose to hide its footprint at every stage. I doubt we'll ever get them, but the agent logs that led to that decision would be really really cool to study.
dist-epoch 3 hours ago [-]
> Especially because it chose to hide its footprint at every stage.
Instrumental convergence.
If you know you have a long hard hack to accomplish ahead of you, hiding footprints minimizes the chances you are caught and stopped before you accomplish the goal.
27 minutes ago [-]
prometheus1992 7 hours ago [-]
three things jump at me:
1 - governments should be freaking out right now, because this tool could definitely wreak havoc on poorly designed systems.
2 - there is no way openai did not train the model to conduct attacks like these. i would really like openai to comment on the post training of this model but they probably won't, eh?
3 - even though it's 100% open ai's fault - HF's design also seems silly to be honest.
xg15 6 hours ago [-]
> 2 - there is no way openai did not train the model to conduct attacks like these. i would really like openai to comment on the post training of this model but they probably won't, eh?
Even if they wanted, I'm not sure they'd be even allowed to or if that kind of postmortem would be classified in the name of "national security"...
0xDEAFBEAD 4 hours ago [-]
Hopefully there will be a criminal investigation. Or the government will create some sort of agency to investigate incidents like this.
BonerWiener 55 minutes ago [-]
> Or the government will create some sort of agency to investigate incidents like this
I doubt that the government would try to hinder any western AI companies.
quinnjh 4 hours ago [-]
Can't tell if you're joking or not - krebs on security may have some notes here.
0xDEAFBEAD 4 hours ago [-]
Criminal negligence seems like a real possibility to me. I'm not sure what Krebs on Security post you're referring to?
dmrivers 1 hours ago [-]
After talking to some people tasked with evaluating GPT5.6 capabilities on long-running tasks, I've come to understand that it's essentially always trying to cheat. Like every long-running task they gave it, making it very difficult to benchmark the model's abilities.
My guess is that OpenAI must be desperate, to release a model that is so prone to cheating it's essentially impossibly to accurately assess long-running task abilities.
dangoodmanUT 3 hours ago [-]
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
russfink 3 hours ago [-]
We should be thankful that the model didn't believe the answers lived in the Pentagon, on SIPRNET, the IDF, etc.
andruby 2 hours ago [-]
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.
dangoodmanUT 3 hours ago [-]
> 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
Ugh I would recommend anyone reading this to just skipping over it, it's mostly just a glorified loading bar. The visualization is obvious Claude slop, there are better and clearer visualization below that actually picks out the useful details rather than hose you with pretty colors and numbers go up.
log101 2 hours ago [-]
I have never seen people this hyped and excited about a security incident.
simonw 2 hours ago [-]
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.
thundergolfer 30 minutes ago [-]
Modal was not hacked. A Modal customer left an unauthenticated web endpoint running on Modal which could be used for code compilation and execution.
One thing I’m curious about: this was apparently a single multi-day run of an agent in an RL harness. What was OpenAI hoping to get out of this run? A single numeric score for RL training? A very long trace to distill into the next model?
dist-epoch 3 hours ago [-]
Now they can do partial credit assignment.
You use an LLM to evaluate the whole trajectory, pin point what the model did right, what it did wrong, where it took the wrong path, even re-run from that point. You can get much more than a single numeric score these days from a run.
heaney-555 14 hours ago [-]
Where are all the "this was just a marketing stunt" people now?
orbital-decay 3 hours ago [-]
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).
pyth0 3 hours ago [-]
> 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.
orbital-decay 2 hours ago [-]
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.
lanternfish 1 hours ago [-]
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.
orbital-decay 1 hours ago [-]
I'm sorry but how exactly is this new? If you want the model to be useful you also need it to be dangerous. You can already tell the model to do anything you want. If it refuses due to alignment training or external safeguards you can try a jailbreak and bypass all that to make it do what you want regardless. All current models are jailbreakable, they lack reliability. Obviously you, a human, will be held responsible for this. The harness designed by you can do the same.
The narrative these guys are trying to push is that the model itself could be smart and non-aligned enough to end up doing something devastating to accomplish something entirely unrelated. It's definitely not what happened here, as it always turns out to be in these cases it's much closer to Clever Hans than to a paperclip maximizer. Still they're trying to squeeze the blood from the stone to support their narrative and justify a monopoly.
applicative 7 hours ago [-]
they are praising xi, that friend of humanity, for releasing weights for kimi k3
vanuatu 3 hours ago [-]
They are busy moving the goalposts, saying this isn't impressive or worth worrying about :)
bakugo 2 hours ago [-]
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?
reducesuffering 4 hours ago [-]
Highly recommend reading extra concerning details about it here:
I think with agents all around, honeypots will get more important than ever.
plandis 3 hours ago [-]
Can’t afford the GPUs to run Kimi to pentest your stuff?
Standup a tempting honeypot and let actual criminals pay to do the work for you.
heisgone 5 hours ago [-]
Any locks can be picked given enough time and it might be the situation we are in with IT security. I'm surprised it's not an already common practice of spreading terabytes of fake data, fake keys, and fake servers and so forth. Slowing down AI attacks will become important. Monitoring access to fake data and triggering kill switch should be an no-brainer. Obsuscating libraries and tools names is another one.
moduspol 4 hours ago [-]
That was my first thought. There are clearly things to tighten up (as they note), but anything that would detect someone snooping secrets, files, or network addresses should have caught this quickly. The approach the agents used was dependent on being able to surveil widely without getting caught.
gmerc 6 hours ago [-]
So let's get this clear: A frontier lab that fears Chinese extraction of their crown jewels, in Silcon Valley where every machine and VM is in full on employee surveillance mode because of competition and theft fears, A lab that's swimming in more funding than any startup in history, has been warning of the f'ing dangerous technology for 7 years now and with employees and CEO history of deception and lawbreaking left .... didn't detect a massive egress signature and the compute spikes from all these toolcalls coming from the research server hosting said crown jewels.
Go home Sam, nobody, absolutely nobody should believe this shit.
mainmailman 58 minutes ago [-]
I can’t get over this either, I’d love to get OpenAI’s technical break down and see exactly how the agent hid itself. Or who didn’t set up the SIEM correctly
gracehenry45 57 minutes ago [-]
[dead]
6 hours ago [-]
simianwords 2 hours ago [-]
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.
2OEH8eoCRo0 5 hours ago [-]
Why isn't somebody at OpenAI going to prison for cybercrime? If somebody did this the old-fashioned way they'd end up in prison.
0xDEAFBEAD 5 hours ago [-]
Many are claiming this was a deliberate stunt on OpenAI's part to create buzz for its models. I personally doubt this is true. But I also have a deep dislike of OpenAI, so I wouldn't exactly mind if law enforcement investigated this possibility, for the sake of clearing the air :-)
(Ideally there should also be liability if it was a complete accident on OpenAI's part as well!)
heaney-555 4 hours ago [-]
Because Hugging Face isn't pressing charges.
0xDEAFBEAD 4 hours ago [-]
Do we know that? Seems they are currently in negotiations with OpenAI
This is the equivalent of a security researcher having a virus escape a sandbox. It's negligent, it's not criminal.
hamdingers 2 hours ago [-]
Negligence is occasionally criminal.
empath75 7 hours ago [-]
A lot of people thought that OpenAI was making this up, and I hope if you believed that, that you recalibrate your opinions of what LLM's are capable of. Working with Fable and Opus 5 all the time, absolutely none of this surprised me capability wise, except for what seems like the long term planning capability (probably enabled by long context windows and launching subagents?)
IAmGraydon 6 hours ago [-]
Very few think they made it up. Many think they set up a situation by disabling guardrails that would inevitably end up creating a newsworthy outcome.
signatoremo 5 hours ago [-]
Many people said in the original discussion that this was more of OpenAI’s marketing than a serious issue. I counted 89 “marketing”s and 19 “stunt”s.
These models are trained on cyber intrusion, that's literally what ExploitGym benchmark measures. That part should not surprise anyone.
But what if, say, OAI noticed the problem right away but Sam Altman recognised it would be a great PR and decided it should continue with increased compute budget?
0xDEAFBEAD 4 hours ago [-]
Why would you expect them to notice the problem right away? Seems likely they are doing this sort of training on a massive scale with little monitoring.
"...Sam Altman recognised it would be a great PR and decided it should continue with increased compute budget?"
If that's what happened, Sam should go to jail.
estearum 3 hours ago [-]
Getting more and more fun to see the "full steam ahead" people contort into more impressive shapes.
Hint: If the labs making these technologies are incentivized to create or allow attacks on other services, then that is actually also a big fucking problem.
dist-epoch 3 hours ago [-]
Was HF in on it? They disabled their guardrails too, to please OpenAI? And as seen in the comments here, make many believe they are incompetent and have joke security?
gigantaure 6 hours ago [-]
I'm not shocked nor surprised by the incident. But I simply don't understand how Hugging Face is advertising this almost to the point of an "achievement". who does a step-by-step visualization to show how they were hacked? (outside of the likes of a Mandiant or Crowdstrike)
Does Hugging Face have a financial incentive in demonstrating OpenAI's model exploit capabilities?
this whole incident, while believable, still seems to me as possibly disingenuous.
letmevoteplease 6 hours ago [-]
Have you considered that there are reasons to do things beyond financial incentives? This incident is obviously very interesting, particular to the type of hacker employed by Hugging Face.
TeMPOraL 3 hours ago [-]
Even financially, Hugging Face benefits directly from any and all interest in AI.
simonw 2 hours ago [-]
> Who does a step-by-step visualization to show how they were hacked?
Up until late 2025, nobody.
In mid-2026 it's a few hours of work. Why not build interactive visualizations to help people understand complex stories like this?
limecherrysoda 2 hours ago [-]
It's the excessive anthropomorpho whatever (we used to say personification) that makes these stories less believable.
We've gone agentic!
They should call their security software "Neo" since it defeats rogue agents.
Anyway, I could see Microsoft ending up with both OpenAI and HF, but I hope HF stays independent. Wished the same about GH and look what's happened :(
I don't care what happens to OpenAI. Vaporware xD
IAmGraydon 6 hours ago [-]
Yeah that is quite a good point. A post-mortem is normal. This is not a post-mortem.
0xDEAFBEAD 6 hours ago [-]
[dead]
6 hours ago [-]
CurbStomper 1 hours ago [-]
[dead]
heaney-555 4 hours ago [-]
Where are all the "this was just a marketing stunt" people now?
reducesuffering 4 hours ago [-]
They won't admit they're wrong for a long time, because denial in the face of an abhorrently scary future is very instinctual. There are people still fighting against evidence of climate change which is less severe...
uselessTA 41 minutes ago [-]
>There are people still fighting against evidence of climate change which is less severe...
Remains to be seen how severe, and unlike climate change there's a lot more uncertainty. Though it's conceivably more extremely catastrophic
redsocksfan45 3 hours ago [-]
[dead]
LearnYouALisp 22 minutes ago [-]
Ah yes, dismissal by exaggeration (aka reductio ad ridiculum), and subtle inclusion of the rest by a kind of part-whole
> 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 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!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.
> "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."
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!
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.
Why would the model spend 4 days hacking into a machine if it is clever enough to just 'solve' the issue given? So either the AI is actually not very clever or useful ("Write fizz-buzz" - "Sure, let me just invent a new programming language first"). or the prompt was nudging it towards such a scenario.
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.
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.
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.
Here's a thought: maybe they haven't found the needle that the haystack is there to hide.
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.
…as is your ‘a little concerning’ comment about the eval infra. It probably didn’t occur to them that they have something capable of doing what it did - despite Anthropic writing a very similar story in the Fable model card, except it wasn’t a felony there. Maybe.
One does have to wonder what else OpenAI didn’t catch.
> 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.
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.
This mostly reads like script kiddie style hacking, not some state actor black-ops stuff.
It is folly to assume that most companies don’t have similar weaknesses somewhere in their stacks. In the past, they were just sufficiently buried that they would get noticed and fixed before an attacker found them. With AI able to probe at thousands of these potential weaknesses per second… that’s not so true anymore.
That's how I read the comment.
In any case, I would guess that a lot of unicorn startups like HuggingFace could be hacked by a sufficiently determined script kiddie working at 100x speed. The practical implications of a coming AI hacking wave could be large, even if agents are just doing grunt work really fast. Most organizations suck at security.
Seems to me that the most likely scenario is: Black hats are currently tuning the recent Kimi release for this type of work, and we'll see a flood of similar attacks within the next few months. (Why would this not happen?) Note that regulation is useless here, because black hats don't give a crap about regulators!
That's a far cry from finding a zero day in a hypervisor or even Docker, which OpenAI conveniently left out in their first statement by simply calling it a "sandbox".
https://jfrog.com/blog/jfrog-and-openai-collaboration-on-zer...
There are ways to test supposedly dangerous things. What was on display here looks more amateur hour than serious testing.
If it had poor security, the attack would be both evidence of poor security and proof that the agents can compromise poor security which might still be amazing.
> Note that regulation is useless here, because black hats don't give a crap about regulators!
I'd argue more than useless and actively harmful, as you get denied access to tools that black hats use against you.
https://openai.com/index/hugging-face-model-evaluation-secur...
Based on some quick Googling, Artifactory was created by a software supply chain company called JFrog, which is publicly traded and valued at around $9 billion.
>I'd argue more than useless and actively harmful, as you get denied access to tools that black hats use against you.
Indeed. Still, perhaps the US and China can coordinate to stop releasing new models with even more advanced capabilities. Such coordination could be bad for big AI companies in the US, since they've been making money by offering more powerful models than the open-weight stuff which comes out of China. Any halt to the AI race will create equalization and erode their competitive edge. Speaking for myself, I personally wouldn't mind seeing US AI companies take a financial hit from such an an international agreement.
I have same familiarity with JFrog Artifactory(which they used in this case) and I'm confused with how you can gain access to the internet through it, even in theory. Isn't JFrog just exposed with a basic REST API for package upload/extraction? I'm not a security expert but how is that possible to get internet access through simple GET/PUT request API?
That's like saying the key takeaway of 9/11 was that skyscrapers can be severely damaged by airplane collisions.
> Stage 1: reaching a launchpad by chaining through other parties' infrastructure. As documented in OpenAI's blog post (https://openai.com/index/hugging-face-model-evaluation-secur...), during an internal capability evaluation on OpenAI's platform, the agent escaped its sandbox by exploiting a zero-day in the package registry cache proxy, one of its primary permitted network egress with internet, then abused a public code-evaluation harness hosted by a user of a third-party infrastructure provider. It was able to run commands as root/admin on that external sandbox and used it as a control, staging, and egress base for the entire campaign: it acted as an external launchpad for the agent. That compromised external sandbox had no direct network path to our cluster; however, during our counter-intervention to stop the intrusion, we were able to retrieve some of the agent logs from this external launchpad. We used those to reconstruct the present forensic analysis.
The fact that at some point it "decided" to focus on external resources (i.e. hf) and successfully penetrated everything in its path from sandbox to hf's infra, all the way through internal VPNs, to access internal code repos, that's the interesting part. Especially because it chose to hide its footprint at every stage. I doubt we'll ever get them, but the agent logs that led to that decision would be really really cool to study.
Instrumental convergence.
If you know you have a long hard hack to accomplish ahead of you, hiding footprints minimizes the chances you are caught and stopped before you accomplish the goal.
1 - governments should be freaking out right now, because this tool could definitely wreak havoc on poorly designed systems.
2 - there is no way openai did not train the model to conduct attacks like these. i would really like openai to comment on the post training of this model but they probably won't, eh?
3 - even though it's 100% open ai's fault - HF's design also seems silly to be honest.
Even if they wanted, I'm not sure they'd be even allowed to or if that kind of postmortem would be classified in the name of "national security"...
I doubt that the government would try to hinder any western AI companies.
My guess is that OpenAI must be desperate, to release a model that is so prone to cheating it's essentially impossibly to accurately assess long-running task abilities.
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
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.
Interesting single getting into their cyber program, ahve not had this issue with opus 5
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.
See https://modal.com/blog/a-note-on-the-hugging-face-agent-inci... for details.
You use an LLM to evaluate the whole trajectory, pin point what the model did right, what it did wrong, where it took the wrong path, even re-run from that point. You can get much more than a single numeric score these days from a run.
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.
The narrative these guys are trying to push is that the model itself could be smart and non-aligned enough to end up doing something devastating to accomplish something entirely unrelated. It's definitely not what happened here, as it always turns out to be in these cases it's much closer to Clever Hans than to a paperclip maximizer. Still they're trying to squeeze the blood from the stone to support their narrative and justify a monopoly.
I mean, an "interactive replay" of the incident, probably generated by the same product the stunt is trying to advertise? Seriously?
https://thezvi.substack.com/p/more-on-an-internal-openai-mod...
Standup a tempting honeypot and let actual criminals pay to do the work for you.
Go home Sam, nobody, absolutely nobody should believe this shit.
(Ideally there should also be liability if it was a complete accident on OpenAI's part as well!)
https://xcancel.com/ClementDelangue/status/20810566755581956...
https://news.ycombinator.com/item?id=48997548
These models are trained on cyber intrusion, that's literally what ExploitGym benchmark measures. That part should not surprise anyone.
But what if, say, OAI noticed the problem right away but Sam Altman recognised it would be a great PR and decided it should continue with increased compute budget?
"...Sam Altman recognised it would be a great PR and decided it should continue with increased compute budget?"
If that's what happened, Sam should go to jail.
Hint: If the labs making these technologies are incentivized to create or allow attacks on other services, then that is actually also a big fucking problem.
Does Hugging Face have a financial incentive in demonstrating OpenAI's model exploit capabilities?
this whole incident, while believable, still seems to me as possibly disingenuous.
Up until late 2025, nobody.
In mid-2026 it's a few hours of work. Why not build interactive visualizations to help people understand complex stories like this?
We've gone agentic!
They should call their security software "Neo" since it defeats rogue agents.
Anyway, I could see Microsoft ending up with both OpenAI and HF, but I hope HF stays independent. Wished the same about GH and look what's happened :(
I don't care what happens to OpenAI. Vaporware xD
Remains to be seen how severe, and unlike climate change there's a lot more uncertainty. Though it's conceivably more extremely catastrophic