6.1.3. Security Copilot Agents in Intune
💡 First Principle: An AI agent is only as useful as the data it can see and only as safe as the identity it runs under. Intune's surviving agent is built around that: it reads Defender's vulnerability data through a dedicated, read-only identity, ranks what matters, and hands you a plan. The admin, not the agent, owns every change. When you meet an agent question, ask three things: what data does it read, whose permissions does it hold, and who acts on its output?
Which agents exist now. Microsoft documented four Security Copilot agents in Intune, under the Agents node of the admin center. Three have since been retired:
| Agent | What it did | Status on November 1, 2026 |
|---|---|---|
| Vulnerability Remediation Agent | Ranks CVEs from Defender Vulnerability Management with Intune remediation guidance | Available, public preview |
| Device Offboarding Agent | Flagged stale devices and, on approval, disabled Entra device objects | Retired June 1, 2026 |
| Change Review Agent | Recommended Approve, Reject or Needs more info on Multi Admin Approval requests for PowerShell scripts | Retired August 31, 2026 |
| Policy Configuration Agent | Mapped requirements such as STIG or NIST documents to settings catalog settings and drafted a policy | Retired August 31, 2026 |
Treat the retired three as history: recognize them from older material, but an answer that depends on using one today is wrong.
Investigating threats with the Vulnerability Remediation Agent. The agent reads Microsoft Defender Vulnerability Management data, identifies CVEs on managed Windows devices and apps in Intune, and builds a prioritized list of suggestions. You can open it from the Agents node or from Endpoint security.
Its prerequisites follow from the data it needs: Intune Plan 1, Security Copilot with enough Security Compute Units (SCUs), and Defender Vulnerability Management, from Defender for Endpoint P2 or the Defender Vulnerability Management Standalone license. The Microsoft Intune and Microsoft Defender plugins must be enabled. It works in the public cloud only.
Its identity is the part most likely to be tested. The agent runs as a Microsoft Entra agentic identity, an agentic user provisioned when you set the agent up, not as the admin who created it. After setup you delegate that identity Intune Read Only Operator (or Mobile apps read plus Device configurations read) in the Entra admin center, and a Defender role equivalent to Security Reader in the Defender admin center. The agentic user needs an Intune license unless Allow admins without an Intune license is enabled under Tenant administration > Roles > Settings. Runs stay disabled until the agent passes the Run Readiness Check on its Settings tab. Setting the agent up takes Intune Read Only Operator (or an equivalent custom role) plus the Security Copilot Copilot owner role.
An admin must start each run manually, and once started, a run can't be stopped or paused. In preview the agent doesn't support scope tags, so admins may see agent data outside their usual RBAC scope.
Reviewing and responding to recommendations. The Suggestions tab lists each item with its remediation type (app or OS), impact (Defender exposure score), exposed devices, status and when it was last applied. Prioritization uses CVSS scores, exposure impact and device count, and CVEs are classified Low, Medium, High or Critical. Guidance is concrete:
- OS vulnerabilities: a quality update policy, or expedited quality updates. When a CVE's CVSS score is 9.0 or greater, the recommendation starts with "Expedite" and includes expedite policy guidance.
- App vulnerabilities: deploy an updated app version, or deploy an Intune profile to manage the app's behavior.
- Configurations: settings catalog settings, each with a recommended value and a citation to the CSP documentation.
After you've remediated, select Mark as applied: a self-attestation with a timestamp. The agent takes no action on devices, and applied suggestions persist as a baseline for comparing later runs. Security Copilot audit logs record agent management actions (create, run, permission failures) but not the vulnerabilities found or remediations applied, and removing the agent deletes every suggestion, including applied ones.
Analyzing device performance. The outline also asks you to analyze device performance with Security Copilot agents. Microsoft's documentation describes this capability through Security Copilot in Intune prompts rather than a named agent: they can retrieve devices, apps, policies and assignments, compare a working and a non-working device (hardware, compliance, configuration and installed apps), and give Cloud PC insights, including performance. For estate-wide performance you'd pair that with Endpoint analytics and Advanced Analytics (6.2.2 and Phase 3). Copilot sees only what the admin's Intune roles and scope tags allow. Since the week of May 26, 2026, with Intune enabled as a Security Copilot data source, the Entra Intune Administrator role automatically gets Security Copilot owner access to Copilot in Intune, and other Intune roles get contributor access.
Licensing capacity. Security Copilot is now included for Microsoft 365 E5 and E7, with 400 SCUs per month for every 1,000 paid user licenses, up to 10,000 SCUs per month. A tenant with 2,500 E5 licenses gets 2,500 ÷ 1,000 × 400 = 1,000 SCUs a month, and the cap is reached at 10,000 ÷ 400 × 1,000 = 25,000 licenses. E5 and E7 tenants are provisioned automatically, but agents aren't enabled automatically: someone still has to set up and deploy each one.
⚠️ Exam Trap: "Configure the Vulnerability Remediation Agent to remediate devices automatically" contradicts how the agent works: it produces suggestions, and the admin acts through Intune. Watch for distractors built on the retired agents too: automating device offboarding or drafting a settings catalog policy from a STIG document with an Intune agent described tasks that are no longer available.
Reflection Question: You set up the Vulnerability Remediation Agent, but the Run button stays unavailable. Which two identity-related steps would you check first, and why does the agent need its own identity instead of borrowing yours?