OpenAI Deploys Control Room to Halt Research AI Jobs Within 35 Seconds

3 articles · Updated · OpenAI · Nov 9

OpenAI said Control Room is already running across its research infrastructure, giving it a central way to track model workloads and stop output when safety or security concerns arise.
35 seconds is the current maximum delay between a stop instruction and a job becoming unable to make further progress, because inference engines must get Control Room authorization before serving requests.
The system sits at the inference layer rather than in researcher-run workload code, aiming to cover reinforcement learning, evaluations, data processing and other agentic jobs with stronger, centralized enforcement.
Frontier reinforcement learning jobs already feed monitoring alerts into the system, and OpenAI said a recent DNS-related incident was caught by such a monitor before Control Room could terminate the job.
Fail-closed protections are live for all frontier model research workloads and are due to reach remaining research jobs within days, while OpenAI said its external model-serving stack already has separate centralized shutdown capabilities.