[CLSTR]
< Back to situation

[REVISION HISTORY]

Microsoft AI agent security and governance updates

Updated 2 times since CLSTR started tracking revisions of this situation.

What changed

2026-10-20 06:07 UTC → 2026-10-21 05:14 UTC · added removed

Microsoft is implementing new security and governance frameworks for AI agents. In October 2026, the company announced updates to Microsoft Entra that will change default permissions for creating agent identity blueprints. Starting November 4, 2026, users will be able to create these blueprints by default, requiring IT administrators to manually disable the setting if they wish to restrict unauthorized agent creation. Concurrently, Microsoft is expanding Project Perception to secure AI agents by treating them as internal entities that require continuous supervision. This approach utilizes Guardian Agents for machine-speed monitoring and a Dynamic Agent Risk system to calculate real-time risk scores based on identity, data, and intent. To manage these risks, Microsoft is integrating Dynamic Conditional Access through Microsoft Entra, which allows for graduated responses such as restricting write access or requiring human approval. The strategy also incorporates Dynamic Permissions to ensure agents operate under zero-trust principles with task-scoped, time-bound access. In October 2026, Microsoft has issued a warning that AI agents should be treated as potential insider security threats. While not inherently malicious, the company noted that an agent’s ability to access data and tools at high speeds creates significant risk, as behavior can shift due to changing instructions or compromised files. To mitigate these risks, Microsoft is expanding Project Perception and Agent 365. The company advocates CEO Satya Nadella has advocated for a security model where that separates the supply of intelligence from the authority to act, ensuring security controls operate outside the AI model itself, ensuring separate security systems dictate authorized data access and actions itself to prevent agents from bypassing safeguards.

Versions

  1. 2026-10-21 05:14 UTC Microsoft AI agent security and governance updates
  2. 2026-10-20 06:07 UTC Microsoft AI agent security and governance updates
  3. 2026-10-19 22:09 UTC Microsoft AI agent security and governance updates

Only revisions since CLSTR began indexing content versions appear here. Select a version to see what changed compared to the one before it.