OpenAI announces Control Room shutdown system for research AI

OpenAI announced Control Room on November 9, a centralized system intended to stop internal research workloads from producing further model output after a safety or security stop order. The company says authorization caching can leave a delay of up to 35 seconds before output is blocked, and it is working to reduce that interval to zero.

Control Room tracks research jobs, including training experiments, evaluations and data processing that use model inference. Each job receives an identifier, which accompanies requests to an inference engine. Before generating output, the engine checks whether the job remains authorized. If a researcher or monitoring system orders a stop, the system denies further requests for that job, even if its research program continues running.

OpenAI says the system is designed to fail closed: if Control Room becomes unavailable, inference requests are denied until it recovers. The company says this protection is live for frontier research workloads and that it planned to extend it to remaining research workloads within days of the announcement. Those rollout statements have not been independently verified in the reports.