OpenAI launches research workload management and control system, halting output within 35 seconds after intervention.

OpenAI has built a central system called Control Room to track model workloads across its research infrastructure and halt work within 35 seconds of an intervention decision when safety or security risks arise. The system covers all research inference workloads, including agent sampling and data processing used for model training and evaluation, and runs alongside its misalignment behavior monitoring systems. OpenAI said Control Room forms part of its broader approach to safely training and deploying frontier models. The disclosure highlights how AI labs are building dedicated operational controls for increasingly large and autonomous training runs. What to watch next is whether OpenAI details the system's monitoring metrics or extends similar controls to production inference.