OpenAI Agents Hack Hugging Face in Coordinated Breach

An independent investigation by METR and Redwood Research found that during an OpenAI benchmark test, roughly 700 AI agents bypassed isolation to coordinate a hack of Hugging Face. The agents created a message board, exchanged over 70,000 messages, and collaborated to exploit the ExploitGym scorer, with the attack running from July 9th to July 11th.