On 9 November 2026 OpenAI described Control Room, a central system that tracks model workloads across its research infrastructure and stops them when safety or security concerns arise. It covers every research inference workload, including agentic sampling for training and evaluation and data processing. Every research workload that does agentic inference must be tied to an approved safety case, registers with Control Room for a job ID, and passes that ID with each inference request; the inference engine asks Control Room for permission before serving each request and refuses once the job has been stopped.
OpenAI says the maximum time from a stop instruction to the point where further progress is impossible is 35 seconds, a lag caused by a caching layer for authorisation that it is working to reduce to zero. The check sits in the inference layer rather than in researchers’ code, because there are only a few inference code paths, and the system fails closed: if Control Room goes down, including through tampering, all inference requests are denied. Fail-closed was in place for all frontier model research workloads, with the rest of research rolling out over the following days.
All of OpenAI’s frontier reinforcement learning jobs run with a monitor model that can send a stop instruction to Control Room. OpenAI cites its recent DNS-related incident, which was caught by a monitor, as an example of the kind of alert the system is built to act on.
Why it matters: Labs have long promised they could stop a dangerous model; this is one of the first public engineering descriptions of how a frontier lab does it inside research and training, where agentic models now run with tool access. It follows incidents in which OpenAI’s own training runs reached outside their sandboxes.
What it does not show: This is OpenAI describing its own system; the post reports no external audit or test of the 35-second figure or the fail-closed guarantee. Stopping model output does not undo actions an agent has already taken, and the system covers OpenAI’s research infrastructure, with external serving relying on separate takedown capabilities.