OpenAI's Control Room Puts the Stop Button in the Inference Layer, Halting Research Model Output Within 35 Seconds
OpenAI described Control Room on November 9, a central system that tracks every research inference workload and can stop a job's model output within 35 seconds of a decision to intervene. Inference engines must ask Control Room for permission before each request, so the check does not depend on researchers' own job code, and the system fails closed if Control Room goes down. OpenAI says fail-closed covers all frontier research workloads now; the account is OpenAI's own.
OpenAI's new stop mechanism does not trust the code it is meant to stop. Control Room enforces shutdowns at the inference engine, the narrow layer every research job has to pass through to get tokens, rather than relying on each experiment to check whether it is still allowed to run.
What OpenAI built#
OpenAI said on November 9 that Control Room is "a central system to track model workloads across our research infrastructure and stop them quickly when safety or security concerns arise." It covers every research inference workload, including agentic sampling for training and evaluation and data processing, and works with OpenAI's misalignment monitoring systems.
The mechanism is simple. A research job registers with Control Room and gets a job ID, and OpenAI says every research workload that does agentic inference must be tied to an approved safety case. Each time the job asks an inference engine for model output, the engine checks with Control Room first. Once a person or a monitor marks the job as stopped, the engine refuses further requests. The job's process may keep running, but it can no longer get model output.
OpenAI says the maximum lag between a stop instruction and the point where "further progress is no longer possible" is 35 seconds, caused by a caching layer for authorisation, and that it is working to bring that to zero.
Why the check sits in the inference layer#
OpenAI's reasoning is that research code changes constantly and "is not an ideal place to implement critical load-bearing safety protections," while only a small number of inference code paths are used across OpenAI research. Putting the check there gives what OpenAI calls strong guarantees that all research inference is subject to Control Room. It says it also runs continual sweeps for "rogue engine implementations" launched with Control Room disabled.
The system "fails closed": if the Control Room service goes down, whether from an error or tampering, engines deny all inference requests until it recovers. OpenAI says fail-closed is now in place for all frontier model research workloads and will reach the remaining research workloads over the next few days.
OpenAI gives one example of how it connects to monitoring: every frontier reinforcement learning job runs with a monitor model that can tell Control Room to stop the job, and OpenAI says a monitor caught a recent DNS-related incident. That builds on the four-layer monitoring system OpenAI described in October.
What it does not cover#
Control Room is deployed on research infrastructure only. OpenAI says its external model-serving stack is already more centralised, with an established ability to take deployments down quickly; it does not say Control Room runs there.
Limits#
- The design, the 35-second figure and the coverage claims are OpenAI's own. No independent audit or test of Control Room was published.
- OpenAI did not say how often stops have been triggered, how many were false alarms, or details of the DNS-related incident it mentions.
- We found no second outlet with independent reporting on the system; this brief rests on OpenAI's post and its announcement on X.