OpenAI has put a switch between its researchers and their models. In a post published this week, the company described Control Room, a central service that tracks model workloads across its research infrastructure and can stop them when safety or security concerns come up.
The system already covers OpenAI's research clusters, the company said, including agentic sampling for training and evaluation, high-volume sampling and data processing. Every research workload that runs agentic inference must now be tied to an approved safety case. The Register called it a kill switch.
How a stop order travels
Each job asks Control Room for a job ID when it starts. When the job needs tokens, the inference engine running the model checks that ID with Control Room before answering. If a person or an automated monitor has marked the job as stopped, the engine refuses the request. The research process can keep running. It just stops getting model output, and the engine stays available to other approved jobs.
OpenAI placed the check in the inference layer on purpose. Research code changes constantly, while only a handful of inference code paths run across the company. That narrow layer is easier to lock down. OpenAI said it also runs automated sweeps to catch engines launched with Control Room missing or disabled.
How fast can OpenAI's Control Room stop a model?
Within 35 seconds. OpenAI says that is the longest gap between Control Room receiving a stop order and the point where a job can no longer get model output, and the delay comes from an authorization cache.
OpenAI said it is working to bring that figure down to zero. The cache exists to handle the volume of authorization traffic.
Built to fail closed
If the Control Room service goes down, for any reason including tampering, engines deny every inference request until it comes back. That fail-closed setting now covers all frontier model research workloads and will reach the rest of OpenAI's research jobs "over the next few days," the post said.
The tool plugs into OpenAI's monitoring. Every frontier reinforcement learning job runs with a monitor, a second model that watches the one being trained and raises alerts. OpenAI said a monitor caught a recent DNS-related incident. Under the new setup, a monitor can send a stop order straight to Control Room.
External products are handled separately. OpenAI said its public model-serving stack is already centralized, with established ways to take deployments down quickly. For teams thinking about how their own AI tools are wired together, our look at the content creator stack covers the consumer end.
Key takeaways
- OpenAI says Control Room is deployed across its research infrastructure.
- A stop order halts a job’s model output within 35 seconds.
- If Control Room goes down, all research inference requests are denied.
What to watch next
Whether OpenAI publishes progress on shrinking the 35-second window, and whether other frontier labs describe comparable controls. Earlier this week OpenAI opened business profiles and agents inside ChatGPT.