Microsoft CEO Satya Nadella says every AI model should be treated as compromised until proven otherwise. In a lengthy Saturday morning post on X, he argued that advanced AI can no longer be handled as "a set of nested black boxes" whose answers and actions users simply accept or reject.
Nadella's proposals, reported by The Verge and TechCrunch, include timely incident disclosure, independent audits, verifiable data and containment. His most quoted line goes further than most industry voices: "We must assume a model is compromised and contain it from the start. Think of it like an emergency brake."
What Nadella wants from AI makers
Nadella framed his post as a call "to step back and assess the trust architecture" of AI. That means separating the model from the harness that orchestrates its work, and externalizing controls and safeguards rather than burying them inside the system.
He also wants every meaningful model action documented with "tamper-proof human readable evidence", so that what a model did can be checked after the fact. On top of that, an authorized person should always be able to pause or shut down a model mid-task, a safeguard he compares to an emergency brake.
- Assume a model is compromised and contain it from the start
- Separate the model from the harness that orchestrates its work
- Externalize controls and safeguards
- Log every meaningful model action with tamper-proof human readable evidence
