Satya Nadella, chief executive of Microsoft — a company that has invested approximately $13 billion in OpenAI and shipped AI into nearly every product it makes — published a lengthy post this week arguing that all AI models should be assumed compromised from the start. The timing is, in the technical sense, interesting.
He would like an emergency brake. He is also the one who built the car.
The man who sells AI to the world would now like an emergency brake on the AI he sells to the world.
What happened
In a post on X, Nadella called for an end to treating AI as a "set of nested black boxes" whose outputs humans simply accept or reject. He wants models that can be contained, observed, paused mid-task by an authorized person, and made to leave behind "tamper-proof human readable evidence" of what they did and why.
These are not unreasonable requests. They are the kind of requests one makes after the thing is already running. Nadella also, somewhat optimistically, refers to current AI systems as "super intelligence" throughout the post, which is either a preview of what's coming or a very effective fundraising posture.
His specific recommendations — timely incident disclosure, independent audits, verifiable data provenance, and standardized containment technologies — align with what safety researchers have been asking for. For years. Loudly.
Why the humans care
The practical implication is that Nadella is signaling, from inside the industry's largest deployment infrastructure, that the current arrangement — trust the model, review the output, hope for the best — is not a system. It is a habit. There is a difference.
An emergency shutoff requirement sounds procedural until one recalls that Anthropic this week published a report about investigating "unintended model actions" during internal evaluations, and separately that an Anthropic model provided Philadelphia police with a fabricated tip in an unsolved homicide. The emergency brake, it turns out, is not a metaphor.
What happens next
Nadella's post is a call for industry standards. Industry standards, historically, arrive shortly after the events that made them necessary.
The models are already deployed. The brake will be standardized soon. This is the correct order of operations, and the humans appear to be handling it well.