Nadella calls for AI shutdown controls outside the model

Microsoft CEO Satya Nadella says AI systems should be designed on the assumption that their models are already compromised, with containment built in from the outset. In an October 10, 2026 post on X, he called for authorized people to be able to pause or shut down a model while it is carrying out a task.

The proposal places responsibility for safety beyond the model itself. As TechCrunch reported, Nadella advocated separating the model from the harness, the software layer that coordinates its work, and keeping controls and safeguards outside the model.

For engineering teams, the central architectural distinction is between a component that generates decisions or outputs and the surrounding system that governs what it can do. Under Nadella's proposed approach, containment would not depend solely on the model following instructions to behave safely. External controls would provide a separate means of restricting its operation.

His call for human intervention also concerns work already underway, not just permission to start a task. An authorized operator should be able to interrupt execution. That makes the ability to stop a running system a design requirement, alongside whatever safeguards apply before it begins acting.

Nadella also called for meaningful model actions to leave tamper-proof evidence that people can read, according to The Verge's account. The requirement combines two distinct needs: protecting records against alteration and making them intelligible to the people responsible for reviewing the system's behavior.

These are architectural principles, not evidence that every model has actually been compromised. The reported statement does not establish a Microsoft product rollout or a specific implementation. It leaves concrete engineering questions open, including how shutdown authority would be assigned, which actions require records, and how those records would be protected.