Local AI agents have quietly become the new shadow IT. Developers install a coding CLI on Friday, and on Monday it reads repositories, runs commands and calls MCP servers with the user’s own access. At Ignite 2026 Microsoft finally gave us a first-party answer: Intune AI agent compliance. A new software compliance policy (public preview) can mark Windows devices running unsanctioned agents as noncompliant. Conditional Access then does the blocking. For approved agents, Intune policies will set boundaries through Microsoft Execution Containers (MXC), but that part is still rolling out.
In this post I cover what was announced, what I can see in my own tenant, and how I would roll it out. One disclaimer: I am writing this on 18 November, and the features are rolling out in phases. The software compliance policy type is not in my tenant yet, and Microsoft Learn has no article for it. I mark clearly what I verified myself.
What Microsoft announced at Ignite 2026
The Ignite 2026 Book of News announces new Intune controls for local AI agents governance, in preview. It has three building blocks:
- Discover agents on managed devices through Microsoft Agent 365, which combines signals from Defender, Entra and Intune.
- Gate devices running unapproved agents. They become noncompliant with Intune software compliance policies, and Entra Conditional Access blocks them from company data.
- Contain approved agents. Intune policies will define what they can access and do inside Microsoft Execution Containers.
Jason Roszak’s What’s New in Microsoft Intune at Ignite post confirms the same story. In addition, the Intune Customer Success team published A guide to governing local AI agents with Intune. That guide holds most of the practical detail.
What I can see in the tenant today
I checked the Intune admin center on 18 November. The software compliance policy type and the Agent containment node are not there yet. Two things are: a new Local AI Agent (Preview) category in the Properties catalog, and the Microsoft Defender AI agent runtime protection Antivirus profile. I cover both below.
How Intune AI agent compliance works
The key thing to understand is that the compliance policy does not detect “AI”. It detects software. You tell Intune which installed apps are restricted, and Intune evaluates the device against that list.
Identifiers: MSI product code or package full name
According to the Customer Success guide, software compliance uses the Platform-specific app ID from Intune app inventory:
- Win32 apps: the MSI product code
- Store apps: the package full name
You add the identifier as a restricted software entry, assign the policy to a device or user group, and pair it with a Conditional Access policy that requires a compliant device. The guide is clear that editor extensions, npm packages and standalone executables can’t be selected in this preview. That limitation matters a lot, and I come back to it under pitfalls. Microsoft has not published licensing requirements for the software compliance policy yet.
What the user experiences
When Intune finds restricted software, the device turns noncompliant. Conditional Access can then block access to protected resources across the entire device, even when the agent is not running. You can write a custom message that explains the reason and the fix. The user uninstalls the software and runs a compliance check. Once Intune confirms compliance, Conditional Access restores access.
Rolling out Intune AI agent compliance: notify first, then enforce
This control blocks access to everything a user needs for work. As a result, I would never switch it on in one step. Here is the sequence I plan to use with customers.
Step 1: Inventory local AI agents first
This step works today. Go to Devices > Configuration > Create > New policy, pick Windows 10 and later and the Properties catalog profile type. Under + Add properties you now find a category called Local AI Agent (Preview). It contains these properties:
- Agent Name, Install Location and Install Scope (mandatory, always selected)
- Agent Version, File Name, Host Process and Publisher
- Install Scope Platform User ID and Install Scope User ID
- Local AI Agent Execution Context and Local AI Agent Installation Method
For a first pilot I select everything except the two user ID fields.

In the same profile, or a second one, add ApplicationProperties with Platform-specific app ID ticked. That column gives you the MSI product code or package name you will need for the compliance policy. In addition, use Agent 365 discovery and the Defender signals if you have the licenses.
Step 2: Agree on the sanctioned list
This is a governance step, not a technical one. Decide with security and the development teams which agents are approved and which are not. Moreover, define an exception process before you enforce anything. If developers learn about the policy from a blocked Teams sign-in, you lose the room.
Step 3: Use the grace period as your “notify” phase
Compliance policies have no report-only mode. However, actions for noncompliance give you a similar effect. Every policy includes Mark device noncompliant at zero days, and you can change that schedule to create a grace period. For Windows, the extra actions are Send email to end user and Add device to retire list. My pilot setup looks like this:
- Send email to end user at day 0, with a notification template that names the restricted agent and links to the approved alternative
- A second email at day 3
- Mark device noncompliant at day 7
Create the notification template first under Endpoint security > Device compliance > Notifications, because the email action needs one.
One caveat: I could only verify this in the classic Windows 10/11 compliance policy. Microsoft’s guide mentions a custom message for the new policy type, but not a grace period. Confirm that the schedule exists before you assign it to anyone outside IT.

Step 4: Put Conditional Access in report-only
Microsoft’s own require device compliance guidance starts the policy in Report-only, and so should you. Scope it to a pilot group and exclude your break-glass accounts. Many tenants already enforce a “require compliant device” policy. In that case report-only does not protect you. The software compliance policy blocks as soon as a device turns noncompliant, so the assignment group and the grace period are your only safety net. Start with an IT-only group.

Step 5: Enforce in rings
The guide recommends piloting on a small group of Windows devices. Validate three things: host-app detection, the resulting compliance state, and how Conditional Access restricts and restores access. After that, widen the assignment ring by ring. I would pair this with the risk-based Conditional Access work from my post on MC1507808 and Microsoft-managed risk remediation, because both change sign-in behavior and you want one communication plan.
Containing approved agents with Microsoft Execution Containers
What MXC is
Microsoft announced MXC as generally available on October 7. The Microsoft Learn overview calls it an open-source, policy-driven execution technology. The agent vendor integrates the MXC SDK, and the workload then runs inside an OS-enforced boundary for files and network. GitHub Copilot and OpenAI Codex already support it. Anthropic Claude Code is listed for later.
The Intune side: agent containment (not in my tenant yet)
The Customer Success guide places the new policy under Endpoint security > Agent containment. You define a “ceiling”, which is the most an MXC-enabled agent can do. Folder access can be deny, read-only or read-write, and network access covers local loopback and external connections. Initial coverage starts with GitHub Copilot CLI and OpenAI Codex CLI.
The status is unclear. The Customer Success guide says preview, while the Windows Developer blog post Enterprise Deployment of Governed Agents on Windows says Private Preview. Plan for a pilot, not production.
That post also describes three modes. Enforcement blocks without a report. Learning blocks and records. Permissive allows and records, and Microsoft recommends it for controlled pilots. In other words, MXC has the report-only phase that compliance policies lack.
Defender runtime protection: visible in the portal today
Compliance gates the device, and MXC contains the agent. Microsoft Defender adds a third layer that inspects what an approved agent does at runtime. This one is already in my tenant: Endpoint security > Antivirus > Create policy > Windows, profile Microsoft Defender AI agent runtime protection.
The profile has three settings under Defender, as the Learn article describes:
- Ai Agent Protection inspects agent activity through vendor-supported agent event interfaces.
- Ai Agent Network Inspection inspects agent-to-LLM traffic for agents without those interfaces.
- Prompt Evidence Collection controls whether prompt evidence is collected.
Each inspection setting offers disabled (the default), block mode, audit mode or Not configured. Note that the UI currently shows the “AI Agent protection is enabled in audit mode” label on the network inspection setting too, so do not let that confuse you. Detections show up as Suspicious AI prompt injection alerts. It requires Defender for Endpoint Plan 2, Microsoft 365 E5, Microsoft Agent 365 or Microsoft 365 E7. Microsoft recommends Audit on a small group for one to two weeks before Block. Status is mixed: the Book of News calls it generally available, while Learn still labels it Preview. Learn lists two known issues. Audit-mode alerts can be titled as blocked, and Claude Code may continue after a reported block. Check your privacy requirements before you enable Prompt Evidence Collection, because it exposes prompt snippets in alerts.

Pitfalls I already see
Coverage gaps in the identifiers
Many agents are not MSI or Store apps. CLIs installed through npm, portable executables and editor extensions cannot be selected in this preview. Consequently, Intune AI agent compliance alone will miss a large share of developer agents. Use the Local AI Agent inventory, Defender runtime protection and Agent 365 discovery to cover that gap.
One agent, several host apps
The guide warns that local agents do not always map one-to-one to host apps. One agent may need several restricted software entries. My own expectation, not from the guide: MSI product codes change with major upgrades, and a package full name contains the version. Review the list every month.
It is access control, not removal
Software compliance does not uninstall anything. It cuts off access until the user fixes the device. Therefore you need clear self-service instructions, or the help desk takes every call. Use a generous grace period.
My take
I like the direction a lot. Intune AI agent compliance reuses the gate we already trust, compliance plus Conditional Access, instead of inventing a new enforcement model. On the other hand, identifier-based detection will never catch everything a developer can run from a terminal.
If you already govern AI spend, as I described in my post on Copilot Credits spending policies, this is the endpoint half of the same conversation. Start the Local AI Agent inventory profile this week. You will need the data whatever the preview ends up looking like.
References
- Ignite 2026 Book of News
- What’s New in Microsoft Intune at Ignite
- A guide to governing local AI agents with Intune
- Enterprise Deployment of Governed Agents on Windows
- Microsoft Execution Containers: Policy-driven containment for AI agents
- Microsoft Execution Containers overview
- App inventory for Windows devices
- Configure actions for noncompliance
- Require device compliance with Conditional Access
- Set up AI agent runtime protection with Microsoft Defender for Endpoint