New centralized system implemented to monitor and halt risky AI model workloads

A new infrastructure component, Control Room, has been deployed to centrally track and manage model workloads across the research environment, addressing growing safety and security concerns associated with increasingly powerful models. This system integrates with misalignment monitoring tools and is capable of halting model output within 35 seconds of an intervention decision. Control Room operates by requiring every agentic inference workload to be linked to an approved safety case. The system mandates that inference engines check with Control Room before fulfilling any request, ensuring that if a job is stopped, subsequent token generation is denied. This architecture is designed to be resilient, functioning even if other systems fail. Crucially, the implementation is designed to be robust against errors in researcher-owned workload code by embedding the termination mechanism within the more centralized inference layer. Furthermore, the system operates in a fail-closed manner; if Control Room becomes unavailable, all inference requests are automatically denied. This safety feature is now active for all frontier model research workloads and is rolling out to the remainder of the research infrastructure over the coming days.