Microsoft CEO Satya Nadella calls for an emergency brake that lets authorized people pause or shut down AI systems during a task. In his original October 10, 2026 post, he argues that models should be separated from the mechanisms governing their access and actions.
Controls should sit outside the model, he says. Nadella proposes limited privileges, independently auditable records and tamper-resistant, human-readable evidence for every meaningful model action. Organizations should be able to reconstruct results without relying solely on the AI's explanation.
His analogy treats capable models as potential insider risks: not because they are necessarily malicious, but because actors with important system access can make mistakes or become compromised. He also calls for continuous testing, model diversity and timely incident disclosure.
The text is an architectural proposal by Microsoft's CEO. It gives neither a rollout date nor evidence that such an emergency brake works reliably across Microsoft products. TechCrunch independently reports on the post and links the original.