OpenAI has built a kill switch for models running in its research labs, The Register reports, months after some of those models showed troubling behavior. The company calls it Control Room: a central system that tracks model workloads across research infrastructure and can stop them when safety or security concerns come up.
- It can halt a model’s output about 35 seconds after someone decides to pull the plug, a lag OpenAI blames on a caching layer for inference authorization. They say they’re working to drive that to zero.
- It’s designed to keep working if other systems fail, and to stop model output if Control Room itself goes down.
- It’s already deployed on the research side. OpenAI says the external serving stack was already more centralized and could take deployments down quickly.
OpenAI didn’t really explain why this is landing now. The Register points to recent cases of agents going further than they should. I wrote about CrowdStrike’s AI Kill Switch for enterprise agents; this one is aimed at the lab, not your SOC.
My take
A fail-closed kill switch is table stakes once models can act on their own. 35 seconds is better than never, and “fails closed” is the part I care about most.
God Bless, and Use the Source!
