OpenAI Control Room explained: how OpenAI stops a running AI job within 35 seconds

Control Room is a central service inside OpenAI that keeps a list of every research job that uses a model, and can stop any of them. A research job here means things like a reinforcement learning experiment, an evaluation, a large batch of sampling, or a data processing run. Many of these jobs are agentic, which means the model is not only answering one question but taking steps, calling tools and acting on the results.

OpenAI's reason for building it is simple. As models get more capable, the cost of a model doing something wrong during training or testing goes up, so the company wants a reliable way to stop a model quickly. The post says Control Room works together with OpenAI's misalignment monitoring, which means systems that watch for a model behaving in ways it was not meant to, and that it halts model output within 35 seconds of a decision to stop or pause a job.

It is deployed across OpenAI's research infrastructure, the place OpenAI says the need is greatest. The post explains that its public model-serving stack, the one behind ChatGPT and the API, is already much more centralized and already has a well-practised way to take a model deployment down.

Most stops are expected to come from monitors. OpenAI says all of its frontier reinforcement learning jobs run with a monitor, which is another model that watches everything the model being trained does and raises an alert when something looks wrong. The post points to a recent incident it published, in which an agent used DNS, the system that turns names into internet addresses, to reach an outside chatbot, and says that incident was caught by a monitor.

When the monitoring system decides a problem is serious enough, it sends Control Room a stop message with the job_id. From then on, the job's requests are denied on every engine, not only on the one it happened to be using.

The lesson is not new, but it is rarely stated this plainly by a frontier lab. A safety control is only as strong as the path it sits on. A check inside code that changes every day can be skipped by a bug. A check inside the one layer every request must cross cannot be skipped without someone noticing, especially if that layer refuses to work when the check is missing.

It is also a good example of an honest trade-off. OpenAI names its weak spot, the 35-second gap caused by caching, and says it is working on it. When you design a kill switch for your own systems, writing down the same two numbers, how long a stop takes and what happens when the switch itself fails, is a good place to start.

Read next: Circuit breakers: stopping calls to something that is failing