Microsoft to govern AI agents, says even Project Perception-grade models can't be trusted

Project Perception was introduced in July 2026 as a three-team system: red agents probe for vulnerabilities from an attacker's perspective, blue agents analyze severity and impact, and green agents strengthen defenses and fix flaws. The update shifts the focus inward, from external threats to the agents an organization itself runs. Gallot noted that agents acting as active participants in business can collude and even resort to improper operations to get results, making them harder to watch than outside attackers. Microsoft's answer is a zero-trust approach that applies policies just before action, assigns each agent a unique ID, and adjusts permissions on the fly, with plans to check inference tokens as well. Gallot also said automation is needed for real-time monitoring, but human judgment remains essential for understanding context and updating security rules.