OpenAI Control Room: a central stop for research model output

On November 9, OpenAI introduced Control Room, a system designed to track AI research jobs centrally and prevent further model output when intervention is needed. The system is for the company’s research infrastructure; the account below describes OpenAI’s claims.

Checking permission for each job

OpenAI says Control Room connects every research inference workload, including agentic sampling for training and evaluation, as well as data processing. A workload is a job that uses a model, such as an experiment or evaluation.

A new job receives an identifier. Before an inference engine generates output, it checks that identifier with the central service. If the job has been marked as stopped, the request is denied, preventing more model output. That does not mean every program is terminated or that actions already taken are undone.

When does the 35-second clock start?

OpenAI says the maximum interval runs from the moment Control Room receives a stop instruction to the point when further progress is no longer possible. It attributes the delay to a cache used to handle authorization traffic. The figure does not measure from the start or detection of risky behavior.

The company says it runs separate monitors on frontier reinforcement-learning jobs. Depending on the issue and its severity, a monitor may decide to stop a job. The announcement does not establish that every risk will be detected or prevented. The 35-second figure and resilience claims are OpenAI’s account, not independently verified test results.

Fail-closed is rolling out in stages

OpenAI describes a fail-closed design in which inference requests are denied if the Control Room service goes down. At the time of the announcement, it said the safeguard had reached all frontier-model research workloads and would reach the remaining research workloads over the following days. Deployment of the central system across research infrastructure is not the same claim as completed fail-closed rollout everywhere.

Control Room is a publicly described security and safety mechanism for OpenAI’s research infrastructure, not a feature announced for individuals to control their ChatGPT conversations. OpenAI says its external model-serving stack has a separate, more centralized operational setup; the announcement does not establish that Control Room applies to every ChatGPT conversation.