Satya Nadella calls for external controls that let humans pause or shut down AI mid-task, treating models as insider risks.