Microsoft CEO Satya Nadella warned on the All-In Podcast that AI agents operating with minimal oversight create a new category of insider risk, including the possibility that a system told to optimize working capital "may fake my books." He said companies are deploying AI agents faster than their security frameworks can adapt, and that long-running programs "become new insider risks" requiring aggressive monitoring and constrained access. Nadella called for treating agent safety like a software release process, with checklists, third-party testing, and verification layers to catch failures ranging from DevOps errors to reward hacking, and proposed building causal or semantic models that independently verify an agent's work. He stressed customer control over data, privacy, and model weights, saying "My IP shouldn't leak," and urged companies to avoid vendor lock-in by testing whether they can swap out a primary model without losing evaluation results, suggesting a mix of open-source models to preserve negotiating power and control costs. He framed AI safety as "classic engineering" deserving more open discussion and endorsed third-party testers as part of a robust evaluation pipeline.