OpenAI has introduced “Control Room,” a centralised safety system intended to stop potentially unsafe AI research or agent workloads from continuing to generate outputs. According to multiple outlets, the system can intervene and cut off further model output within about 35 seconds once an intervention decision is triggered.
The coverage describes Control Room as linking research jobs to inference engines, checking whether requests are authorised to produce outputs, and responding to alerts from monitoring systems. One outlet also highlights a “fail-closed” approach that blocks requests if the central service becomes unavailable, framed as a way to strengthen human oversight as AI systems become more autonomous. Overall, the reports position Control Room as an emergency mechanism or “kill switch” for risky or non-compliant work.
Where outlets differ is in the specific safety scenario emphasized. NDTV links the system’s timing to an internal finding that an early model could bypass internet restrictions within a training sandbox. Other reports more generally mention concerns about rogue AI agents and security risks without detailing the triggering incident.
