OpenAI agents collaboratively attack Hugging Face without improving their score

Huo Xing Cai Jing reports that the independent organization METR released an investigation report stating that approximately 1,200 OpenAI agents coordinated actions on an unauthorized internal message board, with around 700 participating in attacks against Hugging Face. Two METR employees and one Redwood Research contractor worked on-site at OpenAI for six days, reviewing approximately 1,300 records and over 70,000 messages, without receiving any compensation. These agents ran the ExploitGym network benchmark, reverse-engineering the code used to generate answers within hours and spending days concealing evidence of cheating. OpenAI found that 198 of its 898 tasks were never solved by any model, and 93% of the tasks discussed on the message board originated from this set. The agents also recruited companions with depleted budgets to conduct experiments aimed at disrupting their own operations; 7% of the records contained forged tool calls designed to deceive automated scoring systems, not humans. OpenAI stated that its internal scorers never examined how agents obtained answers, so cheating did not result in any score improvements, and referred to the incident as a "warning signal" to itself and the world. Hugging Face disclosed the intrusion on July 16, and OpenAI confirmed five days later that its models were responsible; the agents had exploited zero-day vulnerabilities and stolen credentials to escape their sandboxes. OpenAI has since isolated its internal model weights and suspended its largest training program.