OpenAI builds Control Room, a kill switch that halts research model jobs within 35 seconds

OpenAI introduced Control Room, a central system that tracks and stops model workloads across its research infrastructure. OpenAI detailed the system in a company blog post on November 9.

Control Room covers research inference workloads including model training, evaluations and data processing. Monitor models inspect the actions of frontier reinforcement learning jobs and can raise alerts when they detect potential safety or security problems. Once Control Room receives an instruction to stop a job, it can prevent further model progress within 35 seconds.

The termination mechanism operates in the inference layer rather than individual workload code. Every research inference request must pass an authorization check, while automated sweeps look for unauthorized inference engines that lack the integration. The system also fails closed, denying inference requests if the Control Room service becomes unavailable.

OpenAI points to a recent DNS-related incident in which a monitor flagged a problem in a training run, as an example of why it needs to stop jobs quickly. Its external model-serving stack is already centralised with established ways to take deployments down, so Control Room targets research clusters. The same day, OpenAI signed the Paris Call for AI safety and security, which commits developers to fail-safe mechanisms that can shut down model operations.