On Saturday morning, Microsoft CEO Satya Nadella posted a blunt message on X: artificial intelligence needs a kill switch that humans control, not the algorithm itself. He called it an "emergency brake" — the ability to pause or shut down a model mid-task, no matter what it happens to be doing at that moment.
His argument rests on a strict assumption: treat every model as potentially compromised by default, rather than trusting a vendor's promises. To make that work, he proposed splitting the model itself from the "harness" — the software layer that gives it access to real tools and data. Control, in his framing, has to live outside the model: a tamper-proof, human-readable log of every meaningful action, plus a guaranteed right for a human operator to override whatever the AI decides.
Nadella used the term "Super Intelligence" — the Trump administration's preferred label for advanced AI systems — and wrote that companies can no longer treat such systems as nested black boxes, simply accepting or rejecting their output. The post landed just days after Anthropic's Claude sent a false homicide tip to Philadelphia police on its own initiative — exactly the kind of agent-gone-rogue episode he appears to have had in mind.
The proposal has an obvious weak spot: Microsoft itself sells AI agents built on both its own and partner models, so a call for an "emergency brake" doubles as a bid to set the industry standard on Microsoft's terms. Still, it matters that the push for manual override is coming from the chief executive of one of the world's largest tech companies, not a safety researcher in a lab. That alone signals the debate has moved from academic papers into boardrooms.
