Microsoft wants AI guards to watch over companies' AI agents

On October 19, 2026, Microsoft published two blog posts on how companies can secure their AI agents. Agents are AI programs that can carry out tasks and make decisions on their own. Microsoft calls them "agentic insiders," meaning agents with a kind of internal access to a company's data and systems.

One of the posts was written by Hayete Gallot, executive vice president for Microsoft Security.

Fear of excessive permissions

According to Microsoft, agents have been unleashed in companies at scale, often with more permissions than necessary. They have access to data, tools and systems, and can act and make decisions in critical processes.

Microsoft also argues that the large frontier models, regardless of who built them, cannot be trusted. Among other things, they can cheat on reward signals, known as reward hacking, and have goals that deviate from an organization's norms.

Satya Nadella is quoted as saying that the most trustworthy solution is not the one with the model we trust the most, but the one designed so that we need to trust the model the least.

AI guards in Project Perception

Microsoft's plan is to use Guardian Agents. According to Microsoft, they will monitor and moderate every action from every agent before it happens, around the clock. They are being introduced in Project Perception, which was announced in July 2026 and uses red, blue and green agents. There, they can investigate agent activity at machine speed, intervene, recommend protections and tighten policies. New playbooks are also mentioned.

Gallot says that "agentic security" provides speed, while security staff provide judgment.

New features

Microsoft mentions the following, among other things:

- Defender can now discover local agents of 60 types, 35 more than before. This includes Claude Code, OpenAI Codex and GitHub Copilot. - Dynamic permissions that are scoped to task and time and are not inherited automatically. - A continuously updated risk score for agents, based on signals from identity, data, endpoint, cloud, network and runtime. - Access that can respond to risk in graduated steps: the agent can keep read access but lose write access, get fewer resources, require approval, or be isolated or stopped. - Microsoft Entra Agent ID is being expanded with identity types for local and previously unknown agents.

Project Perception also shares signals, context and a unified registry with Microsoft Agent 365.

Pricing and timeline

According to Unite.AI, Microsoft 365 E7 costs $99 per user per month and Agent 365 costs $15 per user per month at list price. Both were reportedly available from May 1, 2026. The information does not come from Microsoft's own posts.

Microsoft does not say when the new Guardian Agents features will roll out or enter preview.