Microsoft CEO Nadella: all AI models should be assumed to have been "compromised"

Microsoft CEO Satya Nadella posted a long post on X saying that models must be assumed to be compromised and contained from the start, like installing an "emergency brake" for AI, and authorized personnel should be able to pause or shut down a model executing tasks at any time. He advocated for building more transparent systems so that models can be contained and observed, leaving "tamper-proof, human-readable evidence," and proposed timely incident disclosure, independent audits, verifiable data, and containment, adding that more advanced models require more advanced containment techniques that should be standardized.