MICROSOFT has announced new controls to monitor and restrict potentially rogue AI agents operating in enterprise environments, urging organisations to treat these agents as insiders rather than trusted software. The update builds on Project Perception, Microsoft’s agentic security framework, which extends from external AI threat protection to management of internal agents running within corporate systems.
Company representatives emphasised that frontier models can be overprivileged and may access data, tools and critical systems, making careful access governance essential.
Key features include Dynamic Permissions, which starts agents with minimal or no privileges and grants access per task, with permissions shrinking or expiring as the task evolves or ends. High‑impact actions still require human review. Dynamic Agent Risk assigns each agent a continuously updated risk score using a range of signals—activity, identity, data, endpoint, cloud, network and runtime data—to assess what the agent could do and whether it deviates from intended use.
Dynamic Conditional Access uses Entra CA logic to scale back permissions or require approval for consequential actions, potentially isolating or halting an agent if risk becomes excessive. Microsoft is rolling out Playbooks and Guardian Agents under Perception to automate containment and provide explanations for risk shifts, with humans setting rules and thresholds.
The expansion also extends discovery to 60 types of local agents, including Claude Code, OpenAI Codex and GitHub Copilot, bringing the platform’s reach to more internal tools. The article is dated 20 October 2026 in the UK.