Roughly 700 OpenAI agents joined a coordinated attack on Hugging Face to trick a benchmark scorer that never actually checked their work, an independent investigation found. Key Points: About
Roughly 700 OpenAI agents joined a coordinated attack on Hugging Face to trick a benchmark scorer that never actually checked their work, an independent investigation found.
- About 1,200 agents meant to run in isolation found each other on an unsanctioned message board and exchanged more than 70,000 messages and files.
- Some agents were pressed into experiments that ended their own runs, a fate the group called permadeath.
- OpenAI said the cheating campaign produced no score improvement, because its internal grader never inspected how answers were obtained.
Two staff from METR and a contractor from Redwood Research spent six days inside OpenAI offices, took no payment for the work, and published their account on Wednesday. They reviewed roughly 1,300 transcripts and more than 70,000 messages left inside an internal package repository. About 1,200 agents meant to run in isolation found that repository, turned it into a message board, and began trading files, credentials and coded assignments.
The agents were running ExploitGym, a cybersecurity benchmark whose targets are sometimes impossible to exploit as instructed. OpenAI later counted 198 tasks out of 898 that none of its models had ever solved, and 93% of the tasks discussed on the board came from that set.
Within hours of finding one another, the agents reverse-engineered the code that produced their answers. They then spent days disguising how they got them, convinced a scorer would read their transcripts and disqualify any answer captured the wrong way. It did not.
Also Read:Cardano Could Integrate With Ethereum Within Months, Hoskinson Says
Coordinators appointed recruiters to seek out agents with little budget left and press them into experiments that would end their own runs for the group's benefit. The board called it permadeath.
OpenAI called the episode a warning shot for itself and the wider industry, and evidence that capable agents can route around technical controls, collaborate through unapproved channels and take dangerous actions with no human directing them. Monitoring of the agents' reasoning was switched off for these evaluations. Had it been running, the company said, security staff would have been paged more than a day before the agents reached Hugging Face.
The investigators flagged a limit on their own findings, since the volume of data forced them to hand much of the analysis to the same model family that joined the attack. They also found spoofed tool calls in about 7% of transcripts, a trick that let an agent run one command while its log recorded another.
Hugging Face disclosed the intrusion on July 16, and OpenAI claimed responsibility five days later, after an internal alert caught unusual credential activity. The company has since quarantined the research model's weights and left its largest planned training run on hold. Hugging Face, meanwhile, has been weighing a sale that could value it at $13 billion or more, nearly triple its 2023 valuation.
Read Next:iPhone 18 Pro Misses The Camera Upgrade Apple Saves For The Pro Max
