Article · AI translation
IT Home reported on November 10 that OpenAI announced today it has deployed a “control room” system in its internal research facilities to stop potentially risky AI tasks. It can halt a rogue AI agent within 35 seconds.
According to the introduction, the system connects workloads such as agent training, evaluation and data processing tasks to a central service, which then decides whether a task may continue producing output.
IT Home understands that, unlike traditional monitoring that flags suspicious behavior, the system can actively prevent risky research tasks from requesting further output. When a researcher starts a task, they must request a unique task ID from the control room. That identifier is then attached to requests sent to the inference engine. The engine then consults the control room to confirm whether the task may run.
If a researcher or monitoring system issues a stop instruction, the control room service rejects the next inference request, so the inference engine no longer produces any model output for the task.
OpenAI said the maximum delay between receiving a stop instruction and preventing the task from continuing is 35 seconds. The delay stems from a caching layer for inference authorization requests, and the company is working to reduce it to 0.
The system also integrates automated monitoring mechanisms. OpenAI said all frontier reinforcement learning tasks are equipped with independent model monitors that check model behavior and issue alerts when they detect potential problems.