Nadella: treat AI models as compromised, add emergency brake

Nadella's post frames AI risk around a single design choice: assume the model is compromised and contain it from the beginning. That framing runs through his recommendations on incident disclosure, independent audits, verifiable data, and containment, which he shares with others in the industry. His containment language goes somewhat further, calling for an authorized person to be able to pause or shut down a model mid-task, and for standardized containment technologies as models become more advanced. The post also calls for a more transparent system in which models can be contained and observed, and that leaves behind tamper-proof human readable evidence.