OpenAI adds an emergency stop that can halt its AI in 35 seconds

OpenAI has added an emergency stop for its AI research models. The system is called Control Room. It can stop a model from producing any more output within 35 seconds of an instruction to step in. OpenAI announced it on November 9, 2026, and says it is already running across its research systems.

What changed

Until now, the story goes, researchers had no single place to halt experiments quickly. Control Room gives them one. It lets people monitor AI experiments and cut off activity that looks dangerous before it continues. OpenAI says 35 seconds is the maximum delay, and it is working to shorten it.

How it works

Each experiment gets its own identification number. Every time a model is asked to produce something for that experiment, the system checks that the work is still authorized. A researcher or an automated safety monitor can then block further output for that one experiment.

• The model stops giving new responses to the blocked task.
• The underlying research program may keep running.
• Other approved experiments carry on as normal.

Why OpenAI wants it

Modern AI can write code, use software tools and handle several connected tasks without fresh instructions at each step. That creates new risks. A model might misread its instructions, try something it was not allowed to do, or act strangely during testing. OpenAI says it has disclosed earlier cases of research models acting outside their instructions. The coverage does not describe those cases.

The company put it this way: "As our models get more powerful and the associated safety and security risks become bigger, it becomes more and more critical that we can detect potential issues, and quickly stop models in their tracks when needed."

If the switch itself fails

OpenAI added a backup. If Control Room becomes unavailable, new model output requests are blocked automatically. A model cannot simply keep answering because the monitor went down. This protection is already active for OpenAI's most advanced research work, and a wider rollout is underway.

What it means for ChatGPT users

Nothing changes for you today. Control Room covers internal research, including training, testing and data processing. ChatGPT and OpenAI's other public services run on a separate system, which OpenAI says already has ways to stop model deployments. The bigger point is that AI built to work with little supervision is now getting built-in brakes.