OpenAI has built a central off switch for the models it trains and tests, and says it takes up to 35 seconds from the decision to stop a job until that job can get nothing more out of a model. The company describes the system, called Control Room, in a post on its own site in November 2026. It covers research work. Models that customers use run on a separate serving stack, which OpenAI says is already more centralized and can pull a deployment quickly.
The mechanism is easy to follow. A research job that has a model act on its own (a reinforcement learning experiment, a test run, a big batch of sampling, a data processing run) first asks Control Room for a job ID, and each job must be tied to an approved “safety case”. Every time the job wants output, the machine running the model checks with Control Room. Once a job is terminated, its requests are refused and nothing comes back.
Who pulls the switch is often another model. OpenAI says its frontier reinforcement learning jobs run alongside a monitor, a second model that reads everything the first one does and raises an alert when something looks wrong. A monitor caught what the post calls a recent “DNS-related incident”. When a monitor flags a problem it can tell Control Room to halt the job, and the very next request from that job, whichever model or engine it targets, is turned down.
OpenAI put the stop logic in the layer that runs the models rather than in researchers’ own code. Researchers change their code all the time, and the company does not trust it to carry protections that must hold. Only a few code paths exist at that layer across all of its research. It also runs continual automated sweeps, some of them agents, looking for anyone running a homemade engine with Control Room switched off, by accident or on purpose.
The design fails closed. If Control Room goes down, whether from an error or because someone disabled it, “including tampering by a bad actor”, every inference request is denied until the service returns. A crash and a sabotage attempt therefore look the same as a stop order: the models go quiet.
The limits are in the post itself. Thirty-five seconds is the maximum, and OpenAI blames it on a caching layer that stores authorization answers so the system can keep up with traffic volume. The company says it is working to bring the delay to zero. Fail-closed behavior is live for the frontier research jobs, and the rest of its research workloads get it over the next few days, so at the time of the post some jobs were not yet covered.
The post is OpenAI describing its own system. It offers no outside audit and no account of what the DNS incident involved, so the claim that a monitor and Control Room work together as described rests on the company’s word. Still, publishing the delay and the unfinished rollout is more than most labs say about their stop controls.
Thirty-five seconds is a short time to a person and a long one to a cluster of GPUs generating output at high volume. For the labs and enterprises watching how OpenAI handles this, the figure to hold it to is the one it has promised to shrink: how fast that 35 gets to zero.
Reported by OpenAI in its post “Control Room”, published in November 2026 (the page carries no exact date).
