In an essay dated October 10, 2026, Microsoft CEO Satya Nadella proposed that an authorized person be able to pause or shut down an AI model mid-task. He argues that controls over access and actions should sit outside the model, so the system’s authority does not depend on the model policing itself. Nadella’s essay frames this as a governance recommendation.
What the proposed brake would do
The emergency brake is a human-authorized way to interrupt a model while it is carrying out a task. Nadella’s proposal applies to frontier models with either closed or open weights. He also calls for separating the model from its harness—the software that orchestrates it—and from the action space that defines what it can do.
The distinction is between a model’s capabilities and the authority granted to it. Under Nadella’s approach, permissions and action limits would be enforced outside the model, rather than left to the model itself.
The brake is one part of a wider control system
Nadella also recommends tamper-proof, human-readable records of meaningful model actions. Outcomes, he argues, should be reproducible without relying on the model to certify its own behavior. For an organization using AI, that means records should make it possible to examine what the system did—not just what the model says it did.
He calls for ongoing tests covering failures, attacks, edge cases and changes to the system, alongside validation independent of the model being evaluated. His other proposed safeguards include using diverse models, containing systems and promptly telling affected parties about incidents, what controls failed and how recurrence might be prevented.
An insider-risk precaution, not an accusation
Nadella’s essay says frontier models should be treated as potential insider risks, not because they are necessarily malicious, but because capable systems can make mistakes or be compromised. “Assume compromised” is a precautionary design premise; it is not a report that a named model has been compromised.
That framing puts the focus on limiting what a model can access and do, even when the model is capable. The same reasoning underpins the proposed brake: an organization should be able to contain a system without depending on the system to decide to stop.
What the essay proposes—and what it leaves to implementers
Nadella’s essay is a governance recommendation, not an announcement of a Microsoft product or an adopted industry standard. It gives no standardized technical design for the brake and names no job title or authorization process for the person who would operate it.
The essay is separate from Microsoft’s MAI code draft, covered in NeoTeo’s earlier article.
