Introducing Control Room: Human control over AI workloads
Also covered by 1 source
What happened
On November 9, 2026, OpenAI deployed Control Room, a centralized system that monitors and authorizes AI model workloads across its research infrastructure. The system can halt model output within 35 seconds after a safety or security intervention is triggered. Control Room covers all research inference activity, including model training and evaluation, and allows both automated alerts and human interventions to stop workloads. The deployment applies to OpenAI's internal research infrastructure, with external model-serving handled separately.
Key details
- Maximum 35-second lag between stop instruction and halting model output
- Control Room covers all research inference activity at OpenAI
- Fail-closed design: if Control Room is disabled, engines deny all inference requests
- Automated monitors and human responders can trigger stop decisions
- Frontier-model research workloads had fail-closed integration by November 9, 2026
Why it matters
OpenAI’s researchers and security teams now have a centralized tool to rapidly cut off AI model output in response to safety or security concerns, reducing the risk of unauthorized or unsafe behavior during research. This change enables faster containment of problematic workloads and ensures that stop decisions are enforced consistently across the infrastructure. Research teams must ensure their workloads are integrated with Control Room, and any failure or tampering with the system results in immediate denial of model output, affecting ongoing research operations.