Microsoft: Your AI Agents Are the New Insider Threat — Guardian Agents and Live Permissions | inicIA

Direct answer

On October 19, 2026, Microsoft announced new security capabilities that treat enterprise AI agents as potential insider threats. The announcement extends Project Perception —launched in July 2026 to defend companies against AI-driven attacks from outside— to the agents already running inside the organization, and introduces four pieces: Guardian Agents (agents that continuously supervise other agents' actions), Dynamic Permissions (permissions scoped to the task and its duration; the agent never automatically inherits the user's permissions), Dynamic Agent Risk (a live risk score with signals from identity, data, endpoints, cloud, network and agent activity) and Dynamic Conditional Access (if risk rises, the system can block writes while keeping read access, require human approval, isolate the runtime, or stop the agent). The controls connect to Microsoft 365 E7 and Entra Agent ID (a distinct identity for each agent), and discovery covers about 60 types of local agents, including Claude Code, OpenAI Codex and GitHub Copilot. Microsoft gave no prices, dates or preview labels: firm details are expected at Microsoft Ignite, November 17–20, 2026. (Sources: Microsoft's official posts summarized by unite.ai, tokenpost and independent analysis; this is a product announcement, not independently measured security results.)

The mindset shift

The checklist: demand this from your agent vendor

  1. A distinct identity for each agent: every agent gets its own credential, separate from the user's. No identity, no audit trail.
  2. Task- and time-scoped permissions: least access —only what the approved task requires, only for as long as it lasts.
  3. Live risk scoring: continuous behavior monitoring, not a quarterly review.
  4. Human approval for high-impact actions: money, sensitive data, irreversible changes.
  5. A single registry visible to IT: knowing which agents exist —including the ones nobody formally deployed (the "shadow agents" employees connect on their own).
  6. A stop button: isolate or stop an agent without shutting down the business.

Related links

Frequently asked questions

What are Microsoft's Guardian Agents?

Security agents that investigate risk changes in other agents and coordinate protective actions (restrict, request approval, isolate) within human-defined boundaries, as part of the Project Perception system.

Is this available now?

Microsoft announced no prices, availability dates or preview labels. Concrete details are expected at Microsoft Ignite (November 17–20, 2026). Treat the announcement as the market's direction, not a product you can buy today.

Which agents can it detect?

According to Microsoft, its endpoint security covers about 60 types of local agents, including Claude Code, OpenAI Codex and GitHub Copilot, with the full list in its Defender documentation.

My company is small — does this apply?

The principle applies regardless of size: least access, human approval on risky actions, and knowing which agents are connected. You can apply it today with internal policies while the products mature.

Take the next step

AI news only matters if you know what to do with it in your business. inicIA helps you turn every market move into concrete decisions for your company. WhatsApp: https://wa.me/17866120930

Chat on WhatsApp

By inicIA