OpenAI describes Control Room, a 'stop switch' that halts research models within 35 seconds
OpenAI says every agentic research job now registers with a central Control Room service, which can cut off a job's model output at most 35 seconds after a stop order and blocks all research inference if the service itself goes down.
OpenAI has published a technical description of Control Room, an internal system that tracks model workloads across its research infrastructure and can stop them when a safety or security concern arises. It does not cover the separate systems that serve ChatGPT and the API to the public.
How it works
- Every research workload that does agentic inference must be tied to an approved safety case and obtain a job ID from Control Room.
- Each time the job asks a model for output, the inference engine checks with Control Room first. If a person or a misalignment monitor has marked the job as stopped, the engine refuses.
- The research process itself may keep running, but it gets no further model output.
OpenAI says it put the check in the inference layer because research code changes constantly. In its words: “We do not rely on the workload code to correctly perform a check of whether or not the job is still running.”
35 seconds and ‘fail closed’
The maximum delay between a stop instruction reaching Control Room and output ending is 35 seconds, caused by an authorization cache; OpenAI says it is working to bring this to zero. The figure does not include the time needed to detect a problem. The system fails closed: if Control Room goes down, including through tampering, all inference requests are denied. OpenAI says this is live for frontier-model research and will reach the remaining research workloads within days.
The description comes from OpenAI itself and has not been independently audited.
Also available in: Саха тылаРусскийEspañol中文Portuguêsالعربية