OpenAI Control Room can halt any research model workload within 35 seconds and fails closed
OpenAI has described Control Room, a central system that tracks model workloads across its research infrastructure and can stop them when a safety or security concern comes up. According to the company, it covers every research inference workload, including agentic sampling for training and evaluation and data processing, and halts model output within 35 seconds of a decision to stop or pause a job. It is deployed across research infrastructure; OpenAI says its external serving stack is already centralised and has its own ways to take deployments down.
The design has three parts. A workload, such as a reinforcement-learning run or an evaluation, requests a job_id from Control Room when it starts, and every research workload doing agentic inference must be linked to an approved safety case. Whenever the workload wants tokens, it sends the job_id with its request to an inference engine. The engine asks Control Room whether that job may continue; if the job has been terminated, the request is refused and no further output is produced.
The stop signal can come from monitoring. All frontier reinforcement-learning jobs run with a monitor, another model that watches what the model in training does and raises alerts. If the monitoring system decides to stop a job, it tells Control Room, and the job's next inference request on any engine is denied. The 35-second worst case comes from a cache for authorisation decisions needed to handle the request volume; OpenAI says it wants to bring that to zero.
Two engineering choices stand out. The check sits in the inference layer rather than in workload code, because workload code is varied and changes often, while there are only a few inference code paths, deployed uniformly. OpenAI also runs sweeps for engines started without the check. And the system fails closed: if Control Room is unavailable, through error, deliberate disabling or tampering, engines deny all requests until it returns. That behaviour is live for frontier research workloads and is being extended to the rest.

Why it matters
Putting the kill switch on the narrowest shared path, and making its absence stop everything, is the same reasoning as a circuit breaker. It turns a policy into a property of the infrastructure.