Nadella calls for emergency brakes on AI models

On 11 October 2026, Microsoft CEO Satya Nadella posted on X that AI models need an "emergency brake" mechanism, letting authorised people pause or shut down a model while it is carrying out a task. He argued for separating the model from the harness that orchestrates its work, placing controls and safeguards outside it, and requiring tamper-proof, human-readable evidence for every meaningful model action. Follow-up reporting the same day said Nadella wrote in a long post that it is no longer acceptable to treat AI as a set of nested black boxes whose suggestions and actions can only be accepted or rejected, and argued for a more transparent system in which models can be isolated and observed and leave tamper-proof human-readable evidence. He said models must be assumed compromised from the outset and isolated accordingly, and that more advanced models need more advanced isolation techniques that should be standardised. The post calls AI superintelligence in several places.

Written by AI from the reports · updated 5 d ago