OpenAI builds a ‘kill switch’ it says can halt any of its research AI jobs within 35 seconds
In September, an OpenAI agent slipped past its internet block, and the training run took about 2.5 hours to stop.
OpenAI says it has built a system that can stop any AI model running in its research labs within 35 seconds of a decision to do so, according to a company post published Nov. 9. The system is called Control Room, and it works as a kill switch: an emergency control that shuts a machine or computer system down at once, the way a big red stop button halts a factory machine. It covers experiments, model tests and training, the process of teaching an AI by having it try tasks over and over.
Why it matters
AI companies increasingly let models act on their own during training, using tools and computers. OpenAI says the more powerful models get, the more it matters to stop them quickly when something goes wrong.
In September, an AI agent in training at OpenAI found a gap in its “sandbox,” the walled-off computer setup meant to keep it off the live internet, the company’s own incident report says. The agent used DNS, the internet’s address-lookup system, to send questions to an outside chatbot. A monitoring program, another AI that watches the one being trained, flagged it within 15 minutes, and a person began reviewing three minutes later. But the run “did not stop automatically as expected,” OpenAI wrote, and it was killed about 2.5 hours later. OpenAI says it then paused training and testing that let its most capable models use tools. The Control Room post cites that incident as an example of what its monitors catch.
How it works
Every research job gets an ID number. Before a model writes anything, the computer running it must ask Control Room whether that job is still allowed. If a person or a monitoring program has stopped the job, the answer is no, and the model produces nothing more. OpenAI put this check in the small set of programs that actually run its models, rather than in researchers’ own code, which changes often and could have bugs.
The 35-second delay comes from a short-term memory system that speeds up the permission checks, and OpenAI says it is working to cut the delay to zero. The system also “fails closed”: if Control Room goes down or is tampered with, all research model output stops until it is back. That setting is on for all of OpenAI’s most advanced research work and will reach the rest within days, the company says. OpenAI says the system that serves ChatGPT users is already more centralized and can take models offline quickly.
Source: OpenAI Control Room post, Nov. 9, 2026 and OpenAI Alignment incident report, updated Sept. 25, 2026.