무슨 일이 있었나
On Oct 19, Microsoft announced new capabilities in Project Perception and Microsoft 365 E7 that treat enterprise AI agents as insider risks. It expands local agent discovery to 60 types (adding Claude Code, OpenAI Codex, GitHub Copilot), extends Entra Agent ID identity types to local/previously-unknown agents found via sensors, introduces Dynamic Agent Risk, Dynamic Conditional Access, and a team of Guardian Agents plus Playbooks that supervise each agent action and adapt policy as risk changes. Sandboxes like MXC and graduated responses (read-only, narrowed scope, human approval, isolate/stop) enforce containment.
왜 중요한가
This is one of the first major-vendor frameworks that treats agents operating inside the enterprise as first-class insiders — giving agents their own identities, dynamic task-scoped permissions, and runtime-triggered containment rather than relying on the model to be its own authority. It moves agent governance from policy documents to execution-time enforcement across a major platform, with broad reach across Microsoft 365, Entra, and Defender.
적용 범위
Microsoft 365/Entra/Defender enterprises with Copilot or third-party coding agents (Claude Code, Codex) should map the agent estate and evaluate Entra Agent ID + Dynamic Conditional Access and Guardian Agents for least-privilege and insider risk. Pilot against high-risk agents before rolling out.