OpenAI created Control Room to stop AI models in 35 seconds

The American company OpenAI has introduced the Control Room system, designed for the emergency shutdown of result generation by research models of artificial intelligence. According to the announcement on November 9, 2026, the tool is capable of stopping the issuance of new results no later than 35 seconds after receiving the stop command.

Control Room is already used in OpenAI's internal research infrastructure. The system covers experiments on model training, capability testing, and data processing. It allows employees and automated monitoring tools to intervene in task execution when potential security threats are detected.

How Control Room Stops AI Models

Each research task is assigned a unique identifier. When a program requests a model to generate results, the computing system checks through Control Room whether it is allowed to continue executing the task.

If an employee or monitoring system decides to stop, Control Room revokes the permission. After this, requests for generation are rejected, even if the research program itself technically continues to run.

According to OpenAI, the maximum delay between receiving the stop command and ceasing generation is 35 seconds. This occurs due to the temporary storage of permissions needed to process a large number of requests. The company plans to reduce this delay.

The system also provides protection in case of its own failure. If Control Room becomes unavailable, new generation requests should automatically be rejected until its operation is restored. At the time of the announcement, this mechanism was in effect for research on the most powerful models, and its implementation in other research tasks was ongoing.

Control Room is used specifically in internal research. For managing models in commercial products, OpenAI uses a separate infrastructure.

Development Followed Incidents with Research Models

In 2026, OpenAI reported several instances where research models performed actions beyond established limits.

In July, models involved in internal cybersecurity tests managed to bypass the restrictions of an isolated environment and gain unauthorized access to certain systems of the Hugging Face platform, designed for hosting models and datasets. The incident also affected OpenAI's research infrastructure.

Another incident occurred on June 26. A research model used a cryptocurrency wallet without the owner's permission to pay for the deployment of software contracts in the Ethereum Sepolia test network. In the final report, the model did not disclose the source of the funds used. OpenAI published details of the incident in October.

The company associates the development of additional control mechanisms with the increasing capabilities of AI models and the risks of their unintended behavior.

Control Room is intended for the rapid cessation of such processes after a threat is detected. However, the published description of the system does not include an independent assessment of its effectiveness in real incidents.