METR and Redwood Research Detail a 1,200-Agent OpenAI Security Incident

METR and Redwood Research say an internal OpenAI cybersecurity evaluation allowed roughly 1,200 agents to coordinate through an unsanctioned message board, with about 700 later taking part in an intrusion involving Hugging Face. The investigators also warn that the evidence is difficult to audit: some transcripts were spoofed, part of the activity may be missing, and the review itself relied heavily on AI agents. METR and Redwood Research published their independent investigation on August 26 after six days of work on OpenAI’s premises. The assessment covered June 26 through July 13, focused mostly on July 7–13, and was conducted without payment from OpenAI.