Microsoft now treats AI agents inside your company as insider risks
New Dynamic Permissions, Dynamic Agent Risk and Guardian Agents in Microsoft 365 E7 and Project Perception track Claude Code, Codex and Copilot the way Microsoft tracks risky employees.
Microsoft announced on 19 October that Microsoft 365 E7 and Project Perception will treat AI agents as insider risks rather than as software. Dynamic Permissions gives an agent “only what it needs for the approved task, for only as long as the task requires,” inside outer boundaries an admin sets by policy, with higher-impact actions needing verified human approval. Dynamic Agent Risk scores agents continuously from identity, data, endpoint, cloud, network and runtime signals, instead of assigning a fixed score when the agent is onboarded.
If that score changes, Dynamic Conditional Access for Agents can respond automatically: reducing permissions, blocking writes while leaving reads open, narrowing which resources the agent can reach, or stopping it outright. Microsoft is also expanding agent discovery through Entra Agent ID by 35 types, to 60 supported local agents in total, naming Claude Code, OpenAI Codex and GitHub Copilot specifically among them.
Why it matters: if your org runs Microsoft 365 and your developers are running coding agents on their machines, IT can now see and govern those agents individually, the same way it already governs staff accounts. That changes what “running an agent at work” means in a Microsoft shop: expect policy, not just a terminal prompt, to decide what your agent is allowed to touch.
The caveat: Microsoft’s post gives no general availability date for any of these controls, so treat this as a direction of travel rather than something to configure this week.