# Microsoft Outlines Live Access Controls for Enterprise AI Agents By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/35873 Published: 2026-10-20T18:57:32.000Z Updated: 2026-10-20T18:57:32.000Z Section: Technology > Three controls would let companies narrow permissions, block writes or require human approval as an agent’s risk changes during a task. Microsoft has outlined a system to adjust AI agents’ access while they work, allowing companies to restrict sensitive actions if an agent’s behavior raises security concerns. Announced Monday, Oct. 19, the approach has three parts: Dynamic Permissions, Dynamic Agent Risk and Dynamic Conditional Access. They connect an agent’s authority to its task and to risk signals that can change during execution. Dynamic Permissions limits access to a task and its duration. Company policy and the sponsoring user’s rights set the upper boundary, so an agent cannot gain authority the user does not have. Actions with greater impact can require verified human approval. Dynamic Agent Risk assesses signals from identity, data, endpoints, cloud, network, runtime and the agent’s activity. That assessment gives access controls a way to respond to what the agent is doing after its work has begun. Dynamic Conditional Access can then reduce an agent’s permissions. Among the available responses are blocking writes while preserving read access, requiring a person to approve an action, isolating the agent’s runtime or stopping it. Those responses address risks that can emerge during an otherwise legitimate task. An agent may make a mistake, encounter untrusted content or change its behavior as its working context shifts. A permission granted at the start of a task therefore need not remain in place throughout it. The controls add specifics to [Microsoft’s earlier plan for continuous oversight of enterprise agents](). That plan called for supervising agents with access to company data and tools and adjusting their permissions as risk changes. Project Perception, the broader security system behind the controls, was introduced July 27, 2026. Its initial focus was defending against external attacks; the new work extends its scope to agents operating inside organizations. Microsoft is also adding support for 35 local agents, bringing the supported total to 60. Claude Code, OpenAI Codex and GitHub Copilot are among the agents covered by its discovery system. ## Links in this article - [Microsoft’s earlier plan for continuous oversight of enterprise agents](https://www.tokenpost.com/news/technology/34968)