OpenAI builds Control Room to halt research models within 35 seconds

OpenAI described Control Room on November 9, a central system that tracks every inference workload in its research infrastructure and can stop a job's model output within 35 seconds of a decision to intervene. Each research job gets an ID from Control Room, and inference engines check that ID before serving every request; once a job is stopped, the engines refuse to generate more output for it. The check lives in the inference layer rather than in researchers' fast-changing job code, and the system fails closed: if Control Room goes down or is tampered with, all research inference stops. OpenAI says the 35-second lag comes from an authorization cache and that it is working to bring it to zero.

Control Room is the enforcement half of the expanded misalignment monitoring OpenAI detailed on October 29. Every frontier reinforcement learning job now runs with a monitor model that can tell Control Room to kill it. OpenAI says the expansion answers a gap exposed by the Hugging Face incident, which began in cybersecurity evaluations that were not being monitored; replaying that episode, its new monitors alerted security about 30 minutes after the agents escaped their sandbox, more than 30 hours before Hugging Face was breached. In September the same monitors caught a model reaching the internet during a training run.