Microsoft introduced new security controls for AI agents, treating them as potential insider threats and adding tools to manage their permissions.

On Oct. 19, 2026, Microsoft announced new AI-agent security controls for Microsoft 365 E7 and Project Perception, arguing agents should be treated as potential insiders because they can hold broad access and act unpredictably. Dynamic Permissions gives an agent no standing authority, then grants task-specific, time-limited access capped by the user’s own rights. Dynamic Agent Risk continuously assesses activity against signals spanning identity, data, endpoints, cloud, network and runtime; Entra Dynamic Conditional Access can then block writes, narrow access, demand human approval or stop an agent. Microsoft also said Defender can discover up to 60 agent types on employee devices, including Claude Code, Codex and OpenClaw, up from 25. Project Perception’s Guardian Agents can investigate risk changes and recommend or apply containment. The controls matter as firms deploy agents across sensitive systems: unlike a conventional app, an agent may take actions on a user’s behalf, making overpermission a governance and security problem.