Microsoft CEO Satya Nadella says AI systems should start with the assumption that their models are compromised. In an October 10 post on X, he called for containment and an “emergency brake” that lets an authorized person pause or shut down a model during a task, according to The Verge.
The position puts responsibility on the software around a model. Nadella wants controls that people can inspect and operate, alongside records of what the model does. For readers evaluating AI agents, that supplies a concrete question to ask a vendor: who can stop the system once work has begun?
Controls outside the model
TechCrunch reports that Nadella called for a reassessment of AI’s “trust architecture.” He argued against treating advanced AI as layers of opaque systems whose recommendations and actions people simply accept or reject.
His proposed approach separates the model from the software that coordinates its work. Controls and safeguards would sit outside the model. Each meaningful model action would leave a record that people can read and that resists tampering.
The Verge also reports calls for prompt incident disclosure, independent audits and verifiable data. Nadella says more advanced models will need more advanced containment technology, with standards for that technology.
These are design requirements in a CEO’s public post. The two reports describe no Microsoft rollout schedule or product demonstration for the proposed stop mechanism. Nadella’s instruction to assume compromise is a starting assumption for containment, rather than a finding that every model has suffered a breach.
