Microsoft is now treating the AI agents companies deploy as potential insider threats — and on Monday it launched a product to watch them. Guardian Agents continuously monitor the AI agents a company uses, extending Project Perception, the agentic cybersecurity platform Microsoft launched earlier this year with red, blue and green team agents built to defend against outside AI hackers.
Chief executive Satya Nadella's thesis is blunt: the AI enemy may be the very agents an enterprise deployed itself, like a rogue human employee. Frontier models, he argues, cannot be trusted — intelligence must be separated from the authority to act on it.
Guardian Agents arrive alongside support for 35 more agent types — 60 in total, including Claude Code and Codex — and expanded Microsoft Entra Agent ID tooling, with Dynamic Permissions and Dynamic Conditional Access for agents.
The design principles read like an employee handbook for software: least-privilege access, human approval for sensitive actions, and isolation or shutdown when risk runs high. As agents move from chatbots to actors inside company systems, Microsoft is betting the security layer — not the model — is where control will live.
