OpenAI has described Control Room, a system that tracks model workloads across its research infrastructure and can stop their output when safety or security concerns arise. The company says the service controls requests at the inference layer, where the model generates output. A job that has been stopped can no longer get authorization to continue producing tokens.
How does Control Room stop a job?
A research job, such as a reinforcement learning experiment, requests an identifier when it starts. When the job needs model output, it sends the identifier to the inference engine. The engine checks authorization with Control Room before fulfilling the request. If the service has recorded a stop for that identifier, the engine denies further inference for the job. OpenAI also says it requires an approved safety case for research workloads that perform agentic inference.
Why does stopping output take up to 35 seconds?
OpenAI says a stop instruction takes at most 35 seconds to make further progress impossible. It attributes the delay to caching inference authorization so the service can handle the volume of traffic. The company says it is working to shorten that delay. The stated interval begins when the service receives a stop instruction; it does not measure how long it takes to detect a problem or decide to intervene.
What happens if Control Room goes down?
The company describes a fail-closed design: configured engines deny inference requests if they cannot reach Control Room. OpenAI says that behavior is already rolled out for frontier model research workloads, with other research workloads still being brought into the same configuration at the time of its writeup.
OpenAI presents Control Room as an internal research system. Its external serving stack already has a separate deployment control capability. The account explains one control over model output; it does not establish that the service detects every risky action on its own.
