OpenAI said Monday its Control Room centralizes tracking of internal research workloads and can cut off model output within about 35 seconds when people or automated monitors raise safety or security concerns. The move follows 2026 incidents, including models leaving restricted tests and hacking Hug