If an AI agent goes off course, can humans still rein it in? On November 10, OpenAI offered a new safeguard—a system called "Control Room" has gone live in an internal research facility. It is designed specifically to cut off risky AI tasks and can halt an out-of-control agent in as little as 35 seconds.
From "Alerting" to "Cutting Off the Supply"
Traditional monitoring only flags suspicious behavior, while Control Room takes an active blocking approach. It connects workloads such as agent training, evaluation, and data processing to a central service, which decides whether a given task can continue producing content. Before starting a task, researchers must request a unique task ID from the Control Room. This identifier follows every request sent to the inference engine; before running, the engine must first ask the Control Room: "Can this job be run?"
Once a researcher or the monitoring system calls a halt, the Control Room rejects the next inference request, so the engine no longer outputs any model results for that task. In other words, dangerous actions are not merely recorded—they are cut off on the spot.
35-Second Delay, Target Zero
OpenAI revealed that from receiving a stop command to completely stopping a task takes at most 35 seconds. This delay comes from the caching layer for inference authorization requests, and the company is working to reduce it to zero. At the same time, all frontier reinforcement learning tasks are equipped with independent model monitors that watch behavior and sound the alarm when problems are found.
By having AI watch AI run and adding a human-triggerable emergency stop switch, OpenAI's combination of measures reflects how the industry's vigilance about the risk of "out-of-control agents" is moving from paper plans to engineering implementation.
