Microsoft adds risk-based controls for rogue AI agents

Microsoft introduced security controls that restrict AI agent permissions as their assessed risk rises, targeting what it calls the "agentic insider" threat. Discovery tools now detect 60 agent types on employee machines, including Claude Code, OpenAI Codex, and GitHub Copilot. Each agent receives a continuously updated Dynamic Agent Risk score drawn from identity, endpoint, cloud, and network signals. As risk climbs, agents lose write access, face human approval requirements, or are isolated entirely. Permissions become temporary and task-scoped, and a new Guardian Agents team within Project Perception can recommend or apply containment automatically.