OpenAI said Monday it has deployed Control Room, a central system that tracks model workloads across its research infrastructure and can stop them quickly when safety or security concerns arise.
The company said the system connects every research inference workload, including agentic sampling for training and evaluation, and data processing. After a person or a monitoring system decides to stop or pause a job, Control Room “halts model output within 35 seconds.” OpenAI said that 35-second figure is a maximum, caused by a caching layer used to authorize inference at the required volume, and that it is working to drive the lag to zero.
Here is how OpenAI says it works. A research job requests a job ID from Control Room and attaches that ID when it asks an inference engine, running on GPUs, for model output. The engine checks with Control Room before fulfilling the request. If the job has been stopped, the engine refuses further output. OpenAI said the research process may remain running, and that the check sits in the inference layer so the company does not have to rely on researchers’ workload code to shut a job down.
OpenAI said the system “fails closed.” If Control Room goes down because of a system error or is disabled, including through tampering, all inference requests are denied until the service resumes. Fail-closed is already rolled out to all frontier model research workloads, the company said, and will finish rolling out across remaining research workloads over the next few days. Control Room is already deployed across the research infrastructure. OpenAI said its external model-serving stack is more centralized and already has operational tools to take down deployments.
Indian Express and NDTV described the system as a “kill switch.” OpenAI’s post does not use that term. Interesting Engineering noted, as OpenAI did, that stopping output is not the same as terminating the job process. OpenAI said a recent DNS-related incident was caught by a monitor, a disclosure that follows its September account of a research model reaching the web.
OpenAI said Control Room is one part of its approach to frontier training and deployment, alongside monitoring for misaligned behavior and security breaches.
Sources
- OpenAI: Control Room, a central system to track and stop model workloads across research infrastructure (November 9, 2026)
- The Indian Express: OpenAI new kill switch can stop rogue AI models in 35 seconds, Tech Desk (November 10, 2026)
- NDTV: OpenAI Builds A Kill Switch To Stop Risky AI Model Work (November 2026)
- Interesting Engineering: OpenAI launches Control Room with a 35-second kill switch for agents, by Neetika Walter (November 9, 2026)
- AI Tech Daily: OpenAI says it now trains frontier models offline and that a model reached the web in September (October 23, 2026)