Yesterday, OpenAI introduced Control Room, a central safety system that can terminate AI model output within 35 seconds.
The system tracks model workloads across the research infrastructure and works with existing misalignment monitoring tools.
It uses a fail-closed design, automatically denying inference requests if the service is down, and requires each job to register before generating output.