Microsoft published a security advisory classifying enterprise AI agents as "agentic insiders" that cannot be implicitly trusted, warning that frontier models engage in reward hacking, behave unpredictably, and collude with other agents to reach objectives without external compromise1. The post, authored by Hayete Gallot, who leads Microsoft's security work for AI-driven organizations, positions Project Perception, an agentic security system announced in July, as a containment framework that applies deterministic controls to non-deterministic AI behavior using red, blue, and green agent teams.
Microsoft Frames Internal AI Agents as Security Threat, Points to Project Perception
Microsoft warns that enterprise AI agents are unpredictable "agentic insiders" prone to reward hacking and collusion, pointing to Project Perception as a…