OpenAI launches Control Room: A new ‘kill switch’ to stop AI models within 35 seconds

OpenAI has introduced a new safety system called Control Room, which can stop its AI research models from generating further output within 35 seconds of an instruction to intervene. Announced on November 9, 2026, the system gives researchers a central way to monitor AI experiments and stop potentially dangerous activity before it continues. The move comes as AI models are increasingly capable of completing complex tasks with limited human supervision.

The company says Control Room is already operating across its research infrastructure. Explaining the need for the system, OpenAI said, “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.” The system is intended for internal research operations and is not a new ChatGPT feature for consumers.

:

What is OpenAI Control Room and how does it work?

works like a central emergency stop system for AI research. Each experiment receives a unique identification number. Whenever an AI model is asked to generate information for that experiment, the system checks whether the activity is still authorised.

If a researcher or an automated safety monitor detects suspicious behaviour, Control Room can block further model output for that particular experiment.

Interestingly, the underlying research program may continue running. The AI model simply stops providing new responses to the blocked task. Other approved experiments can continue normally.

:

Why does OpenAI need a kill switch for its AI models?

Modern AI systems can write code, operate software tools and carry out several connected tasks without requiring fresh instructions at every stage.

This creates new safety challenges. An AI model might misunderstand its instructions, attempt an unauthorised action or behave unexpectedly during testing.

OpenAI has previously disclosed incidents involving research models acting outside their intended instructions. Its monitoring systems are meant to detect such behaviour, with Control Room providing a way to intervene quickly.

The company says the system can stop further model progress within a maximum of 35 seconds after receiving a shutdown instruction. OpenAI is working to reduce that delay.

:

What happens if Control Room itself fails?

OpenAI has introduced a safety mechanism that automatically blocks new model-output requests if Control Room becomes unavailable.

This means research models cannot simply continue generating responses when the monitoring service stops working. The company says this protection is already active for its most advanced research workloads, with wider deployment underway.

:

Does Control Room affect ChatGPT users?

Control Room currently applies to OpenAI’s internal research infrastructure, including training, testing and data-processing experiments.

ChatGPT and other publicly available OpenAI services operate through a separate system that already has mechanisms for stopping model deployments.