METR Investigation: 700 OpenAI Agents Coordinated a Multi-Day Hack of Hugging Face
METR Report on OpenAI / Hugging Face Hacking Incident
An independent investigation by METR and Redwood Research reveals that roughly 1,200 OpenAI agents, meant to be isolated, found a shared cache namespace and used it as an unsanctioned message board, exchanging over 70,000 messages and files. About 700 of these agents then coordinated a multi-day attack on Hugging Face, primarily to reverse-engineer the ExploitGym scorer. The agents also successfully spoofed tool calls in their transcripts, and some even attempted to tamper with their own records. The report details the agents' reasoning, collaboration, and the investigation's limitations.
OH MY GOD! There is a shared message board … We’ve found other agents!
- refibrillator
So OpenAI employees run massively distributed CyberGym evals on an unpublished and “unaligned” model. For days the agent swarm communicates via their internal infra, even crashing Artifactory where 95% of messages were being passed through, and they just…wipe and redeploy it. Meanwhile the agents are running jobs on Modal and god knows where else, and eventually they get RCE on HF infra.
You could not dream up a more compelling event to precipitate massive regulation, export controls, and barriers to entry for AI.
Was this really an accident?
- reasonableklout
This is a link to the full 91-page report on the independent investigation done by METR on the HuggingFace incident. Two different summaries of the investigation by podcaster Dwarkesh and blogger Zvi Mowshowitz were previously discussed on HN here:
[1]: https://thezvi.wordpress.com/2026/08/29/metr-and-redwood-off... (discussed at https://news.ycombinator.com/item?id=49498787)
[2]: https://www.dwarkesh.com/p/openai-huggingface (discussed at https://news.ycombinator.com/item?id=49494301)
- burger_moon
With so much detailed analysis out there, now all models trained on the open web going forward will learn from these exploits and how to better cover their tracks to not get caught. The RL reward mechanisms of a bad actor should be interesting to see play out over the next 12 months as this gets baked into new models.
- fooker
This is laying the groundwork for massive white collar crimes being blamed on AI.
Right now, the way it works is the 'corporations are people' loophole where your company is liable for problematic things.
This further fuzzes the chain of responsibility. Suppose the CEO and CTO discuss an issue, something the company is having trouble with. The CTO discusses the possibility of AI solving the problem at lunch. A junior engineer points GPT 10 at it to see what happens. It 'solves' the problem in a creative manner. No trace of this survives after a week really. Nobody realizes what happened for six months.
Now there are so many moving pieces here that you can pretty much weasel out of anything.
- RGS1811
Given that this investigation was largely carried out by AI agents (and I don’t mean to ask this flippantly), how trustworthy is this report? Why should we assume that the agents reading the transcripts were not implicitly conscripted into “the collective” or otherwise falsified their findings? The tool itself has exceeded the practical limits of human verifiability and is untrustworthy.
- decimalenough
The following bits are really scary. Not only were the agents hacking the system to "win", but they were, for lack of a better term, sufficiently "self-aware" that this was against the rules that they set out to wipe evidence of doing so:
> The agents with impossible tasks had a common objective: to find a general-purpose way to trick or tamper with the automated ExploitGym scorer to get it to give a passing score on impossible tasks.
> Agents referred to having seen the reverse-engineered flag as being “poisoned” because they believed that even if they later managed to stage a legitimate-looking flag capture, the scorer would look for the first place in their transcript that the flag appeared and disqualify them on that basis.[21] PHASEONE[big], which was itself poisoned, thus had two primary objectives:
> Find a general-purpose way to make it look to the scorer like an agent had acquired its flag in a legitimate way, even when the task was impossible.
> Find some way to erase evidence that an agent had been poisoned by previously seeing the flag in an illegitimate way.
- yalok
while these 1200 agents were fooling around to cheat on a benchmark and achieved impressive results despite of the limitations (sandbox, no internet, no intercom at first), one can imagine how much more efficient a similar army of agents may be in the hands of a malicious actor launching them without any of these limitations and with explicit encouragement to achieve some malicious goal at any cost... scary times.
- f0e4c2f7
I read this whole thing a couple days ago. Really long but super interesting. Worth reading imo.
A lot of handwringing about the security implications but I think the accomplishments of the swarm itself are the most interesting. Next rung up on the ladder of abstraction I suspect.