OpenAI deploys control room system; rogue AI agents halted within 35 seconds

OpenAI has deployed a control room system in its internal research facilities. The system can halt rogue AI agents from continuing to generate output within 35 seconds of receiving a stop command. To operate, tasks must first request a unique task ID from the control room and attach it to the inference request; the control room confirms whether the task can run, and upon stopping, rejects the next inference request. Delays stem from the caching layer for inference authorization requests, which the company is working to reduce to zero. All cutting-edge reinforcement learning tasks are also equipped with independent model monitors that detect potential issues and issue alerts.