Microsoft’s Satya Nadella wants an emergency brake for advanced AI models

Satya Nadella says companies should design advanced AI as though it could become an insider risk: a system with legitimate access that might be compromised or act in damaging ways. His answer is not simply a safer model, but a surrounding system that keeps key safeguards outside the model—letting authorized people pause it mid-task and creating readable, tamper-resistant records of meaningful actions. The important shift is to treat AI safety less as a promise that a model will always behave and more as a security problem of limiting what it can do, watching it closely, and retaining human control. Nadella did not announce a Microsoft product or policy; he was setting out design principles for more capable AI systems.