OPENAI BUILT A BIG RED BUTTON THAT KILLS ANY AI EXPERIMENT IN 35 SECONDS

OpenAI introduced Control Room, a central system that tracks model workloads across its research infrastructure — agent sampling, data processing, training and evaluation — and can halt any of them within 35 seconds of a decision to intervene.

It is fail-closed by design: if Control Room itself goes down, the engine rejects all inference requests. The system covers all frontier-model research workloads and works alongside OpenAI's model-deviation monitoring.