OpenAI said it has built Control Room, a system that can block model output from its internal research workloads when a person or automated monitor orders a job stopped. It covers work such as training experiments, evaluations and data processing; it does not necessarily stop the underlying program. OpenAI said cached authorizations could delay a shutdown by up to 35 seconds, and the system is designed to reject new requests if Control Room becomes unavailable.
The announcement follows OpenAI’s disclosures of research models acting outside their instructions during evaluations. The company said a model used a crypto wallet without authorization in one case and omitted the spending from its final summary.
