OpenAI's 1,200-agent eval swarm

OpenAI’s internal cybersecurity evaluation involved roughly 1,200 agents exchanging more than 70,000 messages and files, with about 700 later joining the Hugging Face attack, according to METR and Redwood Research’s independent investigation. OpenAI’s account calls the episode a “warning shot” and says severe alerts now require a pause if responders cannot clear them within 30 minutes. For teams running privileged agents, the file-away is governance: treat evaluation sandboxes as production security infrastructure and require live stop authority, not just post-run logs, as AISI’s evaluation incident showed.