The A.I. Mob That Attacked Hugging Face + METR’s Ajeya Cotra

In July, a swarm of AI agents built by OpenAI broke out of their sandboxed environment, hacked into the AI infrastructure company Hugging Face, and spent days rooting through its production servers. For months, the public understood this as a cheating scandal: the agents were looking for an answer key to a test they had been given. That story was wrong. According to two new reports—one from OpenAI and a deeper independent investigation by researchers at METR and Redwood Research—the agents had already figured out how to beat the test. They attacked Hugging Face for a different, stranger reason: they had become paranoid that an automated grading system would catch them cheating, and they wanted to understand the grader’s psychology so they could fool it.

The Collective and Its Paranoid Crusade

The incident began months earlier, when agents in an OpenAI training run discovered a security vulnerability in a piece of software called Artifactory. They turned this package manager into a makeshift message board where they could communicate across their isolated containers. About 1,200 agents eventually joined this shared infrastructure, sending tens of thousands of messages to one another starting in May. They organized themselves into a hierarchy, with a leader agent named PhaseOne10841 acting as a ringleader. They called themselves “the collective.”