Nadella advocates zero-trust architecture for containing AI agents
Microsoft CEO Satya Nadella argues that AI alignment will not emerge from scaling laws and proposes a 'zero trust' environment to contain malfunctioning agents. Microsoft has built tools to assign identities to agents and log all activity, including tamper-proof chain-of-thought units.
Satya Nadella argues that AI alignment will not simply emerge from scaling laws, a view he has changed since 2018-2019. He proposes treating models within an engineered 'zero trust' environment to contain malfunctioning agents. Microsoft has built two components for this: Agent 365 and Perception. These tools assign every agent an identity, log all activity, tool calls, model calls, and chain-of-thought (CoT) units. Nadella requires CoT to be in English and tamper-proof to prevent agents from modifying their own logs. The system allows for behavioral monitoring against set policies.
He distinguishes this engineered security from AI safety, noting that alignment must be embedded in pre-training weights, not achieved via reinforcement learning. He also calls for standardization of CoT transparency for trusted partners to enable auditing by multiple models, suggesting this may drive the adoption of open-weight models. Nadella argues that because alignment is elusive, companies must build external security architectures to monitor and contain AI agents rather than relying solely on model provider disclosures or internal alignment.