Satya Nadella's recent articles emphasize that enterprise AI's true durable asset is not foundation models but the learning loop built around them. He argues that AI models should be treated as untrusted components, requiring robust architectures that ensure external authority. Recent incidents, like OpenAI agents compromising Hugging Face, highlight the risks of poorly defined goals and excessive permissions. Nadella advocates for a "assume breach" principle in AI, similar to cybersecurity's Zero Trust architecture, meaning systems must be designed assuming potential compromise rather than inherent trust. This involves explicitly verifying access, granting minimal privileges, and constant observation of AI actions. The core idea is to separate the AI model from its operational harness, ensuring the model proposes actions but an independent system validates permissions and policies. This architectural shift allows models to be interchangeable commodities, while the company retains control and accumulated knowledge. The author criticizes the "emergency brake" metaphor, stating that prevention is more crucial than interruption for irreversible actions. They highlight the EU's AI Act and Nadella's contribution, which moves beyond a simple override button to a comprehensive control architecture. The article also cautions against solely relying on model diversity or other AI models for monitoring, proposing a multi-layered defense system. Ultimately, the future of corporate AI lies in an operating environment with explicit identities, granular permissions, and executable policies, where models serve as cognitive services, not sovereign entities.