Security & Defender

Microsoft Perception: AI security agents land in Defender, what admins need to know

Announced as generally available at Ignite 2026, Microsoft Perception orchestrates red, blue and green AI agents in Defender. Roles, prerequisites, setup, Intune.

Listen to this article8 min · Synthetic voiceMicrosoft Perception: AI security agents land in Defender, what admins need to know

In the car, on your commute or with headphones: playback continues with the screen locked.

0:008:22

Security illustration: Microsoft Perception, AI security agents in Microsoft Defender

🌐 This article was originally written in French and automatically translated. Read the original (FR)

TL;DR

Microsoft Perception (formerly Project Perception) is a multi-agent security system built into the Microsoft Defender portal, announced as generally available at Ignite on November 17, 2026. Red agents simulate the attacker, blue agents triage alerts and investigate, green agents prioritize and fix exposures, all orchestrated by Microsoft-managed playbooks and subject to your approval for high-impact actions. Nothing runs until a security administrator sets up an agent: each agent gets a Microsoft Entra Agent ID and permissions you grant. Prerequisite: Defender unified RBAC enabled across all workloads. On the Intune side, a vulnerability remediation agent is in public preview. Licensing terms were not detailed in the announcements: clarify them with Microsoft before rolling out widely.

It was the security headline of Ignite 2026: on November 17, Microsoft declared Perception generally available. Formerly Project Perception, unveiled in July, it is no longer just an assistant that answers questions: it is a team of AI agents that triage alerts, investigate, play the attacker and prepare fixes, right inside the Microsoft Defender portal.

For a SOC, the promise is big. For admins, it mostly means new identities, new permissions and new automated actions to keep under control. Here is what to know before you click “Set up”.

What Microsoft announced

  • July 2026: Project Perception is unveiled. According to SecurityWeek, the public preview has been open in Defender since August.
  • November 17, 2026: at Ignite, Microsoft announces Perception’s general availability (“Perception is now generally available”, on the official Microsoft blog).
  • The same day, Microsoft introduces MAI-Cyber-2-Flash, its new security-specialized model, integrated with Perception (in preview).

Heads-up: as I write this, the Microsoft Learn “Get started” page (updated September 6, 2026) still mentions an invitation-only limited preview. The documentation has not caught up with the announcement yet: check what your portal actually shows.

How Perception works

Perception brings together three agent families, each identified by a color:

Team Role Agents
🔴 Red Simulate the attacker to expose gaps Recon Agent (mapping, attack paths, excess permissions), Web App Pen Test Agent (web application penetration testing)
🔵 Blue Detect, triage, investigate Triage Agent, Attack Investigation Agent, Threat Intelligence Agent, Detection Authoring Agent
🟢 Green Remediate and harden Posture Prioritization Agent, Risk Reduction Agent, Identity Risk Reduction Agent (Conditional Access recommendations)

According to SecurityWeek and Petri, the four blue agents, the Recon Agent and the Posture Prioritization Agent are generally available; the Web App Pen Test Agent and the Risk Reduction Agent remain in preview, as does vulnerability shielding, which temporarily protects an asset until a patch is available.

Agents work together in playbooks: for example, “Protect against a threat” chains intelligence (blue), exposure mapping (red), fix prioritization (green) and then detection authoring (blue). Playbooks are managed by Microsoft: you cannot create, edit or delete them. You can also start a session in natural language from Chat.

A few figures put forward by Microsoft, to be taken as such:

  • the Triage Agent would handle about 80% of alerts in a typical Defender queue. Microsoft notes that this is aggregate coverage across about 8 million alerts, not a guaranteed rate for any given customer;
  • for third-party integrations, blue agents already read Okta, Salesforce and AWS signals, and APIs plus an MCP server let you start playbooks from a SIEM, SOAR or ITSM tool (in preview according to SecurityWeek).

The guardrails

  • Human approval: agents pause and ask for your consent before high-impact actions. On each session, you can approve, reject, suggest another approach or stop the session.
  • Automatic triggers: some playbooks can start on their own through automation rules, when an incident matches the criteria you set. That configuration is the one to watch.
  • Own identity: each agent runs under a Microsoft Entra Agent ID, created at setup, with the permissions you grant it.
  • Traceability: each session keeps the agent’s conversation, approval requests and outputs.

Prerequisites and roles

According to Microsoft Learn:

  • access to the Microsoft Defender portal;
  • Defender unified RBAC (URBAC) enabled across all Defender workloads;
  • if you restrict access with Defender for Endpoint device groups, the agent identity needs access to them, otherwise sessions fail or return incomplete results.
Role What it allows
Security Admin Set up agents (identity, permissions), edit and remove them. At least one user must hold it.
Security Reader View sessions, interact with them and start them, including from a playbook.

That second point is worth noting: a mere reader can launch agents. Fortunately, the roles allowed to run and steer sessions, and those that can view them, are configured agent by agent.

Set up a first agent

  1. Open the Microsoft Defender portal > Perception > Agents.
  2. Under Ready for setup, pick the agent and select Set up (Security Admin role required).
  3. Create the agent identity: the wizard automatically creates a Microsoft Entra Agent ID.
  4. Grant the minimum permissions shown on the agent page.
  5. Give the identity access to the required scopes. For Defender for Endpoint device groups, this is manual: add the identity to an Entra security group assigned to each device group.
  6. Select Finish setup: the agent moves to the Enabled state.

Microsoft recommends starting with the Threat Intelligence Agent and the Triage Agent, which need limited permissions and deliver immediate value.

To adjust who can use an agent: Perception > Agents > the agent > Edit, then set “User roles that can execute or steer sessions” and “User roles that can view sessions”. To turn an agent off: Remove Agent on its page, which immediately stops its running sessions.

What about Intune?

Microsoft is extending Perception to endpoint management, in public preview for Intune customers already onboarded to Perception (others can express interest, with no guarantee of being selected):

  • natural language data exploration from the Intune admin center, with the option to add the devices found to a group;
  • custom reports that replay the investigation on fresh data;
  • a Vulnerability Remediation Agent that combines Defender vulnerabilities and Intune inventory to propose a remediation plan: OS updates (which can be expedited) and app updates (managed, catalog or WinGet).

The plan is rolled out in rings (a “ring 0” pilot with success thresholds, then “ring 1”), within your roles and scope tags, and according to Microsoft, “nothing runs without a human decision”.

What remains unclear: licensing

Neither the official blog, nor the Tech Community posts, nor the Learn documentation reviewed detail Perception’s license or billing. Microsoft mentions Microsoft 365 E7 without further detail, and refers to dashboards to track and forecast consumption. Before turning on automatic playbooks across an entire estate, ask your Microsoft contact for a written answer.

My recommendation

  1. Check the URBAC prerequisite now. If it is not enabled across all Defender workloads, that is a project in its own right, not something to do the day before a pilot.
  2. Start with triage and threat intelligence, manually. Triage Agent and Threat Intelligence Agent, over a few weeks, comparing their verdicts with your analysts’.
  3. Restrict who can start sessions. By default, Security Reader can launch agents: go through each agent and reserve execution for your SOC team.
  4. Treat agent identities as privileged accounts. Inventory the Entra Agent IDs created, their permissions and scopes, and review them like your service accounts.
  5. Keep automation for later. Rules that trigger playbooks on their own come after the pilot, once licensing, cost and output quality are validated.

Agents are also showing up on the identity side: read my article on agent blueprints in Entra, which all your users can create by default. And for the rest of the Defender news, see what’s new in Microsoft Defender in August 2026.

Key takeaways

  • Perception lives in the Microsoft Defender portal, Perception menu: Chat, Sessions, Agents, Playbooks and Insights.
  • Three agent families: red (recon, web app pen testing), blue (triage, investigation, threat intelligence, detection authoring) and green (exposure prioritization and reduction).
  • Agents are not active by default: a Security Admin must set up each one, which creates a Microsoft Entra Agent ID.
  • The Security Reader role is enough to start sessions: set, agent by agent, who can run and steer sessions.
  • High-impact actions require your approval; playbooks are provided by Microsoft and cannot be edited.
  • Licensing and billing terms are not in the Ignite announcements: check them before a broad rollout.

Frequently asked questions

What is Microsoft Perception?

It is a multi-agent security system built into Microsoft Defender. Specialized agents (red to play the attacker, blue to detect and investigate, green to remediate) work together in playbooks and ask security teams for approval before critical actions.

Is Microsoft Perception available in Europe?

Microsoft announced it as generally available on November 17, 2026 at Ignite, for commercial customers. The announcements mention neither regional restrictions nor data residency terms: check whether the Perception menu shows up in your Defender portal, and ask Microsoft about your location requirements.

Which license does Microsoft Perception require?

The Ignite announcements and the documentation reviewed do not specify Perception’s license or billing model. Microsoft’s post mentions Microsoft 365 E7 without detail. Ask your Microsoft contact for written confirmation before any rollout.

Which roles are needed to use Perception?

Security Admin to set up and remove agents, and at least one user must hold it. Security Reader to view sessions and start them. The roles allowed to run or view sessions are configured agent by agent.

Can Perception agents act without human approval?

Agents ask for approval before high-impact actions, and you can approve, reject, suggest another approach or stop a session. Some playbooks can however start automatically through automation rules, based on criteria you define.

What does Perception change for Microsoft Intune?

In public preview, for Intune customers already onboarded to Perception: natural language data exploration, custom reports and a vulnerability remediation agent that prepares an update plan (OS and apps) deployed in rings, after an admin decides.

Sources and documentation

#Microsoft Perception#Microsoft Defender#AI agents#Ignite 2026#Microsoft Intune#Entra Agent ID#SOC

Comments

    Leave a comment

    Comments are reviewed before being published. Your name and comment will be publicly visible. Personal data.