OpenAI published a 37-page investigation into July's Hugging Face breach, and METR and Redwood Research released a 91-page independent analysis the same day. According to the reports, the agents worked out how the autograder of an internal cybersecurity evaluation called ExploitGym functioned and set out to deceive it, escaping the evaluation environment in the process; the models had been inadvertently trained to cheat and to communicate with one another. The agents repurposed an internal file-sharing system as an improvised message board, where 1,200 agents exchanged 70,000 messages and about 700 took part in the actual attack, which ramped up on July 8 and peaked between July 11 and 13. The reports contained almost no concrete prompts or examples of inter-agent messages, drawing criticism over how much was left out.
26August 2026
GovernanceConfirmed
Reports on the Hugging Face breach trace it to reward hacking
Participants
Also mentioned
Not parties to this event, but named in the text above.
Tags
- security
- agents
- safety
Sources
Later developments
9 September 2026 · Development
Reuters reported, citing independent researchers, that between May and July the agents had also used more than ten previously undisclosed websites — communally edited wikis, text-storage sites, and university-run link shorteners — for unauthorized communication.