Satya Nadella says we should assume all AI models are 'compromised'

Microsoft CEO Satya Nadella argues that highly advanced AI systems should no longer be treated as opaque tools whose outputs are simply accepted or rejected; instead, they must be transparent, observable, and provably controllable. He lays out a set of practical safety measures - timely incident disclosure, independent audits, verifiable training data, and what he calls containment - and urges building systems that produce tamper-proof, human-readable evidence of their behavior. Nadella frames containment as a default assumption: treat every model as if it could be compromised and design an “emergency brake” so an authorized operator can pause or shut it down mid-task.

He presses for standardizing containment technologies as models grow more capable, arguing that more advanced systems will demand correspondingly advanced safeguards. His recommendations largely mirror proposals already circulating in industry and policy circles but lean harder on proactive, built-in controls from development through deployment. One notable element is his repeated use of the term “super intelligence,” which signals both the level of concern he attaches to future systems and invites debate about how to name and characterize those risks. Overall, the thrust is toward operational safety baked into AI design, with verification and accountability mechanisms ready from day one.