Back to news
policyOpenAI2026-11-09

OpenAI deployed a Control Room system on November 9 that can halt a runaway AI agent within 35 seconds

The system routes agent training, evaluation and data processing loads through a central service that decides whether a task may keep producing output, and after a stop instruction the service refuses the next inference request, with a maximum delay of 35 seconds that the company is working to bring to zero

On November 10, OpenAI announced it has deployed a Control Room system in its internal research facilities to stop AI tasks that may pose risk, capable of halting a runaway AI agent within 35 seconds.

The system routes loads such as agent training, evaluation and data processing jobs through a central service that then decides whether a task may keep producing output. Unlike conventional monitoring that only flags suspicious behavior, it can actively stop risky research tasks from requesting further output. When a researcher enables a task, the researcher requests a unique task identifier from the Control Room, and that identifier is attached to requests sent to the inference engine, which then asks the Control Room whether the task may run. If a researcher or a monitoring system issues a stop instruction, the Control Room service refuses the next inference request so the engine no longer claims any model output for the task.

OpenAI said the longest delay between receiving a stop instruction and halting further task progress is 35 seconds. That delay comes from a caching layer on inference authorization requests, and the company is working to reduce it to zero. The system also integrates automated monitoring, with every frontier reinforcement learning task equipped with an independent model monitor that checks model behavior and raises alerts when it detects potential problems.

Control RoomOpenAI智能体安全模型监控强化学习