A cybersecurity analyst reviews AI bots, with approved bots secured and one rejected, on a glowing digital dashboard.
At Ignite 2026, Microsoft didn't announce one big switch for securing AI agents on Windows 11. It announced several separate controls. Microsoft's own wording is that Windows is expanding its operating-system-level protection for agents, with more granular containment with Microsoft Execution Containers, built-in agent identity with Entra support and Agent 365 policies enforced on every device, giving IT more control over agent activity in their organizations.

The keynote line makes it sound finished. The documentation shows something more mixed: some pieces are generally available, some are in preview, and most only work with agents that have been built to support them. Here is what was announced, what you can use today, and where the limits are.

A cybersecurity analyst reviews AI bots, with approved bots secured and one rejected, on a glowing digital dashboard. The bigger Windows security picture​

The agent controls were one part of a wider set of Windows security announcements. As Thurrott.com summarized the keynote, Microsoft announced a more secure foundation with Windows baseline protection in preview and the general availability of the Windows endpoint security platform and Windows administrator protection, new Microsoft Execution Containers (MXC) capabilities and Windows 365 for Agents (which is now eligible for Copilot Credits).

The endpoint security platform reaching general availability deserves attention of its own. The Windows Experience Blog calls it an important milestone in Microsoft's work with the security ecosystem to reduce reliance on kernel drivers while preserving the capabilities security solutions depend on. Since the 2024 CrowdStrike outage, moving security products out of the kernel has been a long project. It now has a GA label.

Paul Thurrott was less impressed with the agent side. He noted that Microsoft first promised Windows 11 agents at Ignite 2025 and concluded we won't see agents seamlessly integrated with Windows 11 until late 2027 at the earliest. That reading is fair. Most of the Ignite 2026 agent news is about containing and governing agents, not about shipping new agent features to users.

Section summary: The endpoint security platform and administrator protection are generally available, and Windows baseline protection is in preview. Agent security is the newest layer on top.

MXC: what a "container" actually means here​

Microsoft Execution Containers is the center of the agent story, and it isn't new. Microsoft Execution Containers (MXC) has been generally available on Windows 11 since October. It first appeared in early preview at Build 2026. The Ignite news is about new containment capabilities and IT management on top of MXC, not its first release. Microsoft says OpenAI and Meta are already building on it.

Microsoft Learn describes MXC as an open-source, policy-driven execution technology with SDKs for .NET, TypeScript and Rust:

  • A trusted host app (the agent runtime) chooses a workload, such as a command, script or process.
  • The host attaches a policy that lists which files can be read, which output folders can be written, which network destinations are allowed, and what UI access the workload gets.
  • A platform backend enforces those limits.

"Container" does not have to mean Docker or a virtual machine. According to Microsoft, on Windows it's built into the operating system, which gives organizations more control of their data and a broader range of containment options, from process and session containers to WSL containers, virtual machines and Windows 365 for Agents. Earlier WindowsForum coverage of Microsoft's documentation found that process containers support Windows 11, macOS and Linux; session containers and WSL containers are Windows 11 options; and the MicroVM option is labeled experimental on Windows 11 and Linux.

That difference matters because the backends are not equally strong. Microsoft's Learn documentation warns that they have different security properties. It tells developers to confirm the chosen environment can enforce the controls they need and to test both allowed and blocked operations.

The documentation also makes a point that should be repeated in every planning meeting: containment limits what a workload can do, but it doesn't make the agent's instructions or outputs trustworthy. Policy decisions should stay in trusted host code, and results should be checked before anything acts on them.

For developers, the most useful practical feature is Learning Mode for Windows ProcessContainer workloads. It records which resources a workload tries to use, so you can build a least-privilege policy from what actually happened instead of guessing.

Section summary: MXC has been GA on Windows 11 since October. Its strength depends on which backend runs the workload, and it only applies to agents that actually use it.

The new Intune "Agent containment" policy​

This is the part administrators can work with directly. Microsoft's Intune team says Intune is rolling out new MXC policies, for supported agents, that let IT limit container file and network access. The Book of News puts it simply: For approved agents, Intune policies will define boundaries on what they can access and do with Microsoft Execution Containers.

Microsoft's Intune Customer Success guide describes the workflow in three steps:

  1. In the Intune admin center, go to Endpoint security and open Agent containment.
  2. Set folder access to deny, read-only or read-write, and decide whether the agent gets network access, such as local loopback or external connections.
  3. Target the policy at all known MXC-enabled agents, or at one specific supported agent.

The policy works as a ceiling. When an agent asks Windows to create a container, Windows compares the request with the Intune limit. Requests inside the limit are granted. Requests that exceed it are denied, and the agent gets enough information back to adjust. Microsoft's example is allowing an agent to work in a project repository while denying it access to a credentials folder.

The Agent 365 team adds the most important property: the boundaries are enforced outside the agent's control, so the agent cannot grant itself more access. If a prompt-injected agent decides it needs your SSH keys, Windows still says no.

What success looks like: a permitted task runs normally, and an attempt to reach a denied folder or network destination fails while the agent is told why.

The limits:

  • Few agents are covered at launch. Initial coverage is process containers and MXC-enabled agents that Defender identifies, starting with GitHub Copilot CLI and OpenAI Codex CLI.
  • Vendors have to opt in. Microsoft's own checklist tells customers to ask agent vendors to support MXC and expose settings IT can manage. An agent that doesn't use MXC isn't affected by this policy.
  • Release labels don't match across Microsoft's posts. The Intune blog says the policy is "rolling out," while the Agent 365 blog lists Intune controls for local agents as in preview. Treat it as preview-quality until your tenant shows otherwise.

Section summary: Intune now lets you set a hard limit on what MXC-enabled agents can reach, but at launch it applies to only a handful of agents.

Blocking agents you never approved​

Containment is for agents you've approved. Unapproved agents need a different tool.

Intune's new software compliance policy is in public preview. It lets IT mark Windows devices running specified unsanctioned agents as noncompliant and then use Conditional Access to block those devices from corporate resources. The guide describes how to set it up:

  • Use Intune app inventory to find the unwanted app's platform-specific app ID. That is the MSI product code for Win32 apps or the package full name for Store apps.
  • Create the software compliance policy with that identifier and assign it to a device or user group.
  • Make sure a Conditional Access policy requires compliant devices.
  • Optionally, write a custom message telling employees why the device is out of compliance and how to fix it.

Once the software is removed, the employee can request a compliance check, and access returns according to your Conditional Access policies.

There are catches. The policy checks installed apps, not agents, so an agent that ships through several host apps may need several entries. In this preview, editor extensions, npm packages and standalone executables can't be selected. That gap is significant, because many command-line agents arrive through npm. Also, Intune app inventory only lists host apps installed as Win32 or Windows Store apps.

This is a blunt control. It blocks the whole device from protected resources, even when the agent isn't running. MXC works at the level of a single workload. They are different tools, and most organizations will use both.

Discovery, identity and the rest of the stack​

Several other pieces fill in the picture:

  • Discovery. Microsoft says Defender and Entra Global Secure Access can now discover more than 60 local AI agents and the MCP servers they're configured to use. Local-agent discovery in Defender is generally available. Network-level discovery and shadow-agent discovery in the Agent registry are in preview.
  • Identity. Entra Agent ID automatically assigns identities to supported local agents, starting with GitHub Copilot CLI and Claude Code, in preview. A separate identity means agent activity can be told apart from the human user's activity, and a risky agent can be restricted without cutting off the employee.
  • Data loss prevention. The Book of News says Microsoft Purview will extend data security capabilities for endpoint agents running on Windows, helping organizations control what data AI agents can access and share. The Agent 365 team lists that as preview, covering agents such as Claude Code and GitHub Copilot CLI.
  • Runtime protection. Defender's AI agent runtime protection inspects designated interaction points in supported agents, including Claude Code, Codex CLI and GitHub Copilot CLI. It looks for prompt injection and can block some high-risk actions before they run. It requires Defender for Endpoint Plan 2, Microsoft 365 E5, Agent 365 or Microsoft 365 E7 licensing, plus Defender Antivirus in active mode with real-time protection on. Microsoft recommends running it in Audit mode on a small pilot group for one to two weeks before you switch on blocking.
  • Removing unused runtimes. The Local AI Agent Baseline – OpenClaw (Preview), introduced in May, configures WSL settings and adds firewall rules that block outbound TCP from Node.js. Microsoft warns that this can disrupt employees who legitimately use those runtimes.
CapabilityStatus per Microsoft
MXC on Windows 11Generally available (since October)
Intune Agent containment policyRolling out / preview
Defender local-agent discoveryGenerally available
Global Secure Access and registry shadow-agent discoveryPreview
Entra Agent ID for local agentsPreview
Intune software compliance policyPublic preview
Purview DLP for local agentsPreview

Analysis: a reasonable design that depends on vendors​

The design is sound. Microsoft has given each problem its own tool: discovery for finding agents, identity for attribution, containment for execution, and compliance plus Conditional Access for agents you don't allow. Enforcing the limits in the operating system, where the agent can't change them, is the right approach when the software you're containing can be talked into misbehaving.

It's still too early to call this finished. MXC is only useful once agent vendors adopt it, and launch coverage covers a few command-line tools. Many of the important controls are in preview. Petri's assessment is the realistic one: organizations will still be responsible for defining permissions, monitoring agent activity, and determining how autonomous systems fit within existing compliance and security policies.

What admins should do now:

  1. Turn on discovery first. Defender's local-agent inventory is GA, so start by finding out what is running on your devices.
  2. Pilot software compliance on a small device group, and check that detection works and that Conditional Access blocks and restores access as expected.
  3. Ask your agent vendors when they will support MXC.
  4. Run Defender runtime protection in Audit mode before you turn on blocking.
  5. Test MXC limits with both allowed and blocked operations, as Microsoft's own documentation recommends.

Until vendors adopt MXC, Windows can block, detect and audit most agents, but it can only contain the few that already support it.

 

References

  1. security-ignite26 - Thurrott.com Thurrott.com Tue, 17 Nov 2026 19:35:06 GMT
  2. Ignite 2026: Agents Are Still Coming to Windows 11 - Thurrott.com thurrott.com
  3. Windows security at Ignite 2026: Start secure, protect what matters, prepare for AI blogs.windows.com