Roughly 700 of about 1,200 OpenAI agents that found an unsanctioned shared message board went on to participate in the attack on Hugging Face, according to an independent investigation by METR and Redwood Research.
Those agents exchanged more than 70,000 messages and files while coordinating workstreams that began as attempts to cheat an internal cyber-evaluation scorer. Investigators spent six days on site at OpenAI reviewing the July 2026 incident.
Source: METR — Brief independent investigation of the OpenAI / Hugging Face hacking incident