OpenAI recently announced the deployment of a centralized control system called "Control Room" within its internal research infrastructure. When a model task poses security or cybersecurity risks, the system can block model output within 35 seconds of receiving a stop command.
