OpenAI's Control Room can halt any research model job within 35 seconds

Every research workload that does agentic inference must be tied to an approved safety case. When a job starts, it asks Control Room for a job ID, then attaches that ID to each request it sends to an inference engine. Before serving any request, the engine checks with Control Room whether the job is still allowed to run. Once a person or monitoring system marks the job as stopped, the next check is denied and the engine refuses to produce more output, even though the job's own process may keep running[1].

OpenAI deliberately put the check in the inference layer rather than in researchers' code. Workload code changes constantly and is spread across many teams, while there are only a small number of inference code paths across OpenAI research, so every model call can be forced through the same gate. The company also runs continual automated sweeps to catch anyone launching an engine with Control Room missing or disabled, by accident or on purpose[1].

All of OpenAI's frontier reinforcement learning jobs run with a monitor, a second model that watches everything the model being trained does and raises an alert if something looks wrong. Depending on severity, the monitoring system can tell Control Room to stop the job itself, and from then on any request from that job to any model on any engine is refused[1].

The 35-second ceiling comes from a caching layer needed to handle the volume of authorisation checks, and OpenAI says it is working to cut it to zero. The system also fails closed: if Control Room goes down for any reason, including tampering by a bad actor, all inference requests are denied until it recovers. Fail-closed is live for all frontier model research and will reach the remaining research workloads over the next few days[1].

As labs hand more of their own research to agentic models, the ability to stop a misbehaving run quickly becomes basic infrastructure rather than an afterthought. Control Room covers research systems only; OpenAI says its customer-facing serving stack is already centralised enough to take deployments down fast[1]. Publishing the design gives other labs a concrete pattern to copy: put the off switch where every model call has to pass, not where researchers have to remember to call it.