OpenAI deploys AI "control room": blocking model output for up to 35 seconds

Abstract:

OpenAI recently announced that it has deployed a central control system called "Control Room" in its internal research infrastructure. When the model task causes safety or network security risks,

the system can block the model from continuing to output for up to 35 seconds after receiving the stop command

According to OpenAI, Control Room can uniformly track research tasks such as reinforcement learning training, model evaluation, agent sampling, and data processing, and is linked to the model behavior monitoring system.

The monitoring model will check the reasoning and operation process of the trained model. Once potential problems are discovered, an alarm can be issued or the task can be suspended directly according to the degree of risk

During the specific execution process, the monitoring system will send the task ID that needs to be stopped to the Control Room. The next time the relevant task applies for model inference, the inference engine will verify its running permission; if the task has been marked as terminated, the system will reject the request to prevent the model from continuing to generate content.

OpenAI stated that

Currently, it takes up to 35 seconds from the time the Control Room receives the stop command until the task cannot continue running at all.

The system also uses a "fail-close" mechanism. If the Control Room fails, is shut down, or is tampered with, all related inference requests will be denied by default until service is restored. This mechanism now covers all OpenAI’s cutting-edge model research tasks and is being expanded to other research tasks. OpenAI said that its external model services adopt a more centralized technical architecture, and currently have corresponding deployment offline and emergency response capabilities.