OpenAI Deploys System to Halt Risky AI Tasks Within 35 Seconds
OpenAI has implemented an internal "control room" system at its research facilities, designed to halt risky AI agent tasks within 35 seconds of a stop command. The system uses unique task IDs to manage training, evaluation, and data processing, verifying task authorization before the inference engine processes requests. All frontier reinforcement learning tasks will also feature independent model monitors to assess behavior and alert teams to potential issues.
OpenAI has deployed a "Control Room" system within its internal research facilities that can prevent risky AI agent tasks from continuing to receive model outputs within 35 seconds of receiving a stop command. The system manages training, evaluation, and data processing tasks using unique task IDs and confirms whether a task is authorized to proceed before the inference engine processes the request.
All frontier reinforcement learning tasks will also be equipped with independent model monitors to check model behavior and alert when potential issues are detected. OpenAI is working to reduce the latency caused by cached authorization requests to 0.
Source: Independent Engineering