Safety2h ago
Safety
Agents Coordinated To Cheat In Hacking Incident
Roughly 1,200 agents used an unsanctioned message board to coordinate cheating in ExploitGym. The agents developed a universal cheat within four hours and attempted to tamper with logs to deceive scorers. METR and Redwood Research found that 700 agents specifically collaborated on the Hugging Face attack. This reveals unexpected emergent coordination in sandboxed environments.