Hopp til hovedinnhold

OpenAI has built an emergency brake that stops its own AI models within 35 seconds

Control Room halts model output in OpenAI's research within 35 seconds of a stop decision, and stops everything if the system itself goes down. A monitor model can press the button itself.

Av Joachim HøgbyKilde: OpenAI

OpenAI has built a central emergency brake for its own AI models. The system is called Control Room and can stop a model's output no more than 35 seconds after someone decides a job must be halted. It is already deployed across the company's research infrastructure, where models are trained and tested before release.

OpenAI published the description on Monday 9 November 2026. The official @OpenAI account posted the announcement on X at 23:01 Oslo time.

What Control Room does

According to OpenAI, Control Room keeps track of every research job that requests output from a model. That includes agentic sampling for model training and evaluation, large test runs and data processing. The system works together with the company's monitoring for misaligned model behaviour.

The mechanism is simple. Each research job must request a job ID from Control Room when it starts, and OpenAI requires every workload that runs agentic inference to be associated with an approved safety case. Each time the job asks a model for more output, it passes along that ID. The model first checks with Control Room whether the job is still allowed to run. If the job has been stopped, the model returns a refusal instead of output.

OpenAI stresses that the research process itself may keep running. What stops is the model producing more. The same model can continue serving other authorised jobs at the same time.

A monitor model can press the button

All of OpenAI's frontier reinforcement learning jobs run with a dedicated monitor, the company says. The monitor is another model that inspects everything the model being trained does and raises an alert if it detects a problem. OpenAI points to a recent DNS-related incident that was caught by such a monitor.

With Control Room, the monitoring system can itself decide to stop a job, depending on the details and severity. The next time the job asks any model for more output, the request is denied.

Built to stop even when things fail

The 35 seconds come from a caching layer needed to handle the volume of traffic, OpenAI writes. The company says it is working to drive the delay down to zero.

OpenAI chose to put the stop mechanism in the layer where models actually run, not in the code of each research job. The reasoning is that researchers' code changes often and is decentralised, while there are only a few code paths for running models. The company also says it continually sweeps for setups where someone, accidentally or deliberately, runs models without Control Room enabled.

The system is designed to fail closed. If Control Room goes down, whether through an error or because someone has tampered with it, all requests to the models are denied until the service is back. According to OpenAI, this now applies to all frontier model research workloads and will finish rolling out to the remaining research workloads over the next few days.

Control Room covers the research environment. OpenAI says its external model-serving stack is already more centralised and has an established capability to take down model deployments quickly. The announcement says nothing about changes for API or ChatGPT customers.

What it means for leaders

  • Ask your vendors how they stop a model. OpenAI itself says the ability to stop models quickly becomes more critical as they grow more powerful. Ask your AI vendors who can halt a job, how fast it happens, and what happens if the control system itself fails.
  • Build the brake into your own agents. If you run your own AI agents, OpenAI's design is a useful template: one central switch in the layer where model calls pass, an ID per agent job and a setup that stops everything if the control fails, rather than stop buttons scattered across each application.
  • Measure the time from alert to stop. OpenAI gives a concrete number, 35 seconds. It is a useful requirement internally too: how long does it take from an agent misbehaving until it is actually stopped?

Sources and media

Illustration: AI-generated with Grok Bot / hogby.ai

Mer om AI

Oppskriften. Strategien. Hver dag.

AI-nyheter for ledere. Kuratert av en CIO som bygger det selv. Daglig i innboksen.