What Is Microsoft's Trust System for AI Agents?
Satya Nadella says Microsoft Agent 365, with perception, is the trust system Microsoft built for AI agents. It gives every agent an identity, logs activity, tool calls, model calls, and chain of thought, keeps the logs tamper-proof so an agent cannot edit them, sets policy, and checks behavior against the policy.
He says models are becoming more capable and scaling laws are working, but an assumption that alignment would just fall out of scaling does not seem to hold, and mechanistic understanding is elusive. The posture he wants is an engineered zero trust environment that contains any malfunctioning agent before it does harm. Agents that are more capable and less aligned need a trust system you can rely on. Alignment, in his take, has to show up in training, maybe even in pre-training, and he does not think you can RL your way to alignment. It has to be embedded in the weights at the beginning. The trust architecture sits outside AI safety and is what Microsoft and cyber teams do. Harry Stebbings asks whether the architecture exists or is only symbolic of the progression. Nadella says the architecture exists. People conflate an engineered security zero trust system for containment with AI safety. AI safety is hard, elusive, and not yet achieved, so he wants the trust system built outside the model, with aggressive policy and monitoring for policy deviation. He also wants chain of thought in English so multiple models can check it.
