OpenAI has introduced Control Room, a safety feature designed to shut down AI agents that go rogue within 35 seconds.
The company said Control Room will connect every research inference workload. That covers agentic sampling used for model training and evaluation, data processing and other tasks. Placing these workloads under one system gives OpenAI a single point from which to oversee how its research models run.
Control Room will also work together with OpenAI's misalignment monitoring systems. Those systems are meant to catch cases where a model's behaviour departs from what its developers intend. Pairing that monitoring with a central shutdown mechanism means a problem flagged in one workload can be acted on quickly.
The focus on agentic sampling is notable. In that kind of workload, models take multi-step actions rather than returning a single response, and those runs are used to train and evaluate the next generation of systems. Oversight of such activity is therefore tied directly to how OpenAI develops its models.
The announcement material did not set out how Control Room detects problems or what triggers a shutdown. It also gave no rollout timeline. The 35-second figure is the headline claim, and the technical basis for it was not explained.
Control Room reflects a wider push among AI developers to build controls around increasingly autonomous systems as agents take on more work. For OpenAI, the tool brings monitoring and intervention together across its research infrastructure.
