AI Agents as Insider Threats: Inside Microsoft's Bet on a Unified Security Stack
Dark Reading, Friday, October 23rd, 2026
Microsoft's Guardian Agents treat AI agents as insiders with their own identities and risk scores; analysts doubt agents can police agents.
Microsoft expanded Project Perception to discover agents, give each a distinct Entra identity starting with no permissions, score agent risk continuously and cut off anomalous agents through Guardian Agents and dynamic conditional access.
Forrester's Andras Cser doubts Microsoft's third-party integrations and calls agents watching agents "non-deterministic nature squared," while Okta argues no single vendor sees enough.
Experts note the worst outcomes can occur inside an agent's authorized permissions.
Suzu Labs advises starting now at no cost by inventorying every agent, assigning unique expiring credentials and requiring human approval for consequential actions.