The guide is useful, but it needs careful reading. Some important parts are still in preview. Several controls work only with certain agents. And "the policy was assigned" does not mean "the policy is being enforced," a lesson the guide repeats throughout.
Where this fits in the MXC timeline
MXC is not new this week. Microsoft on June 2, 2026 announced an early preview of Microsoft Execution Containers, a cross-platform SDK meant to contain AI agents on Windows and WSL. In October it shipped: Microsoft Execution Containers reached general availability on October 7, 2026, sandboxing AI agents on Windows 11, macOS and Linux.
At GA, central management was still missing. Microsoft's October post said Intune policy will soon be available to manage MXC process containers used by MXC-integrated agents on Windows 11. These policies will offer IT administrators control over how Windows evaluates container creation requests made by agents, and the resource boundaries that are enforced by those containers. 4sysops summed up the result: until Intune and Entra attribution ship, the boundary on a PC is the policy the agent vendor shipped.
The new guide moves that forward, though not to general availability. It labels its Intune walkthrough "Configure MXC with Intune (Private Preview)." Most tenants should treat it as a planning document, not a switch they can turn on today.
Section summary: MXC itself is GA. Intune-managed MXC policy is in private preview. The new guide is a playbook for the full enterprise stack, and parts of that stack are still arriving.
The three foundations: identity, containment, manageability
Microsoft's argument is that traditional application control was never designed to follow an agent across tools, child processes, files, network endpoints and long-running sessions. It names three OS-level foundations:
- Identity. When an agent acts for a user, its actions should be distinguishable from the user's own. Microsoft Entra Agent ID covers the identity side, and Windows provides local isolation boundaries.
- Containment. An agent should reach only what its task needs. Microsoft's example is an agent building a website: it should see the project folder, not your personal files. MXC enforces this through platform controls, not policy guidance alone.
- Manageability. Intune or another MDM distributes policy. Entra governs authentication and access. Agent 365, Defender and Purview add observability, security and data protection.
Microsoft calls identity "the control anchor." Note the qualifier that appears throughout the guide: support varies by agent and scenario. Microsoft's Entra documentation describes an agent identity as a special service principal with no credentials of its own. Credentials live on the agent identity blueprint it was created from. Agent identities are always single-tenant, even when their blueprint is multitenant. None of this means every local Windows agent process automatically carries an Entra agent identity. The integration has to exist.
Which product does what
| Layer | Product | What it does in this design |
|---|---|---|
| Identity and access | Microsoft Entra Agent ID | Separates agent activity from user activity; controls access to cloud resources |
| Execution containment | MXC | Runtime limits on file reads and writes, callable resources, and allowed endpoints and MCP servers |
| Policy distribution | Intune or other MDM | Delivers MXC policy, device compliance, app protection, blocks unauthorized agents, turns on Defender runtime protection |
| Threat detection | Microsoft Defender | Local agent inventory, runtime protection against prompt injection, alerts, investigation |
| Data protection | Microsoft Purview | Endpoint DLP for supported agent prompts, tool calls and tool output |
| Estate governance | Microsoft Agent 365 | Agent registry, MCP connection visibility, Dynamic Agent Risk, agent templates |
The key point is that these layers complement each other; none replaces another. MXC limits where a workload can go. Microsoft's own MXC documentation warns that containment does not make an agent's instructions, outputs or permitted actions trustworthy. A contained agent can still be manipulated through prompt injection into doing something harmful inside its allowed space. That is the job of Defender and Purview.
What "container" means in MXC
The name sounds like Docker, but MXC is not Docker-style containerization. Microsoft Learn says an MXC workload can run in an OS process sandbox, a separate session, or a virtualized environment. The available features depend on the MXC release, the backend and the host.
In the MXC v1.0 V1 SDKs, Windows supports three backends: ProcessContainer, IsolationSession and WSLc. Windows Sandbox, MicroVM and Hyperlight are experimental and can't be created through the V1 SDK APIs. TechAIWire describes the session container as running under a separate Windows account and session, with its own desktop, clipboard and input.
That range matters for rollout planning. The guide says a coding assistant may need only lightweight process isolation, while a fully autonomous agent may need a separate Windows session. Microsoft's platform-support documentation adds two cautions:
- Runtime discovery and probing show what a host supports, but they don't validate your whole policy.
- A Windows version can be supported by MXC and still lack a particular policy feature.
The only reliable check is a negative test on each target configuration. Have the workload try to write outside its allowed path and confirm that the attempt fails.
Step 1: Inventory first
The guide starts with something less exciting than containment: an inventory. Record which agents are running, under which identities and credentials, who uses them, and for what work. Also record which tools and plug-ins they call, which files they touch, and which network destinations they contact. Microsoft suggests the Agent 365 Registry as a starting point.
For each workload, the guide asks for four views:
- Business objective: the result the user needs and why the agent needs access.
- Execution model: how the work flows across processes, tools, commands, files and endpoints.
- Control boundaries: where identity context, data sensitivity, network exposure or admin privilege changes.
- Operating readiness: OS and agent versions, management settings, security policy, deployment groups, support and governance.
Microsoft also points out a common mistake: treating policy as the whole deployment. A successful rollout also needs version eligibility, tenant configuration, security-group design, communications, support routing, rollback planning and change approval.
Defender's local agent inventory is another useful source, with its own requirements. Microsoft Learn lists:
- Defender for Endpoint Plan 2
- Onboarded devices
- Defender Antivirus in active mode with real-time protection
- On Windows, antivirus platform 4.18.26090.9 (2609) or later
Risk levels, risk indicators and recommendations need Microsoft 365 E7, or Agent 365 plus Defender for Endpoint Plan 2. Microsoft also notes that seeing an agent's process can be enough to create an inventory entry. An entry does not prove anyone actually used the agent.
Step 2: Learning Mode, carefully
Writing a least-privilege policy is hard when you don't know what the workload touches. MXC addresses this with three operating states:
| State | Behavior | Intended use |
|---|---|---|
| Enforcement | Ungranted access blocked, no activity report | Production |
| Learning (enforced + recorded) | Ungranted access blocked and recorded | Diagnosing denials, confirming only required access |
| Permissive (allowed + recorded) | Access the policy would deny is allowed, but recorded | Controlled pilots |
Microsoft's Learning Mode documentation adds several cautions:
- Allow-and-record runs the workload with the user's full permissions in effect. Use it only with trusted code and non-sensitive data on a test device or account.
- Learning Mode works only with Windows ProcessContainer workloads. The guide says the feature is available through the MXC SDK.
- Capture can change the isolation tier. On some hosts, "guarded capture" starts an elevated helper and requires a UAC prompt or admin credentials. The captured run may therefore differ from production, so finish with a run that has all capture settings removed.
- Results are evidence, not approval. Capture output helps author file, registry, clipboard and UI policy. Some recorded activity doesn't map to a policy setting, and some failures aren't logged. An empty report does not prove the workload is safe.
The guide's advice: exercise first-run setup, authentication, updates, error recovery and rarely used workflows, not just the expected path. Then review the results with the workload owner and security team, and remove incidental access before you write rules.
Step 3: Write the minimum policy and roll out in rings
Turn the validated activity into explicit grants for specific folders, domains and tools. Keep the policy modular, so one risky capability can be changed without reopening everything else. Record a rationale, owner, review date and rollback plan for every exception.
The ring structure will look familiar to anyone who has run Windows Update for Business:
- Ring 0, technical validation: the deployment team proves assignment, version checks, telemetry, exclusion and rollback.
- Ring 1, informed early adopters: users close to the workload test real tasks and the support documentation.
- Ring 2, limited production: more users across typical device and network conditions.
- Ring 3, broad deployment: only after success criteria are met and support capacity is in place.
Set promotion criteria before each ring starts: installation and assignment success, task completion, performance, alert volume, help-desk contacts, a confirmed rollback, and no open high-severity security issues.
For critical workloads, Microsoft recommends a predefined break-glass fallback, such as a temporary switch to permissive mode. It should be limited in duration, explicitly approved, fully logged and reviewed afterward, with a named owner responsible for returning to enforcement. Without those conditions, a "temporary exception" tends to become permanent.
Test opt-out and rollback in Ring 0. Confirm that removing a user or device from the assignment actually restores the previous settings. Document whether groups are user- or device-based, who controls membership, and how long changes take to propagate.
Step 4: Configuring MXC in Intune (private preview)
For tenants in the preview, Microsoft describes this workflow:
- Define the scope: all MXC-capable agents, or one named agent such as GitHub Copilot CLI.
- Treat Intune policy as the upper limit. Agents may let users narrow access further, but only within IT's limits.
- Create an MXC containment policy, scoped to a specific agent if needed.
- Set file system boundaries: read/write or read-only access where needed, including the agent's own install path. Deny sensitive locations such as corporate OneDrive folders, Program Files and core system files.
- Set network and runtime controls: allowed and blocked destinations, and whether permissive mode is allowed for approved diagnostic or pilot scenarios.
- Assign to pilot groups first, then later rings.
- Check behavior: inside the write scope, the agent can modify files. Outside it, the agent can read but not modify. On denied paths, access is blocked.
Before assigning, confirm that targeted devices run an MXC-capable Windows build and the supported agent version. Then check the actual MXC behavior on representative endpoints, not just the "succeeded" status in the console. Microsoft also sets a useful production gate: administrators must be able to create, assign and validate the policy without help from engineering.
Step 5: Turn on Defender, starting in Audit
Defender's AI agent runtime protection is in preview. It targets prompt injection at the device level. There are two inspection methods, each with Disabled, Audit and Block modes:
Set-MpPreference -AiAgentProtection <mode>for agent-native event inspectionSet-MpPreference -AiAgentNetworkInspection <mode>for network inspection
In the stored values, 0 = Disabled, 1 = Block, 2 = Audit. Note that Block is 1, not 2.
Intune deployment: create an endpoint security Antivirus policy for Windows and configure the Ai Agent Protection setting. That profile covers agent-native event inspection only. For network inspection, Microsoft says to use a PowerShell platform script. Set "Run this script using the logged on credentials" to No so it runs in the system context.
Check the result: in the Defender portal, open the device page and go to Configuration management > Effective settings. Close and reopen any terminals used to run agents, then run Microsoft's benign prompt-injection demonstration.
Microsoft recommends one to two weeks in Audit on a small device group before moving to Block. Two preview issues to know:
- An Audit-mode alert may incorrectly report a block.
- In Block mode, Claude Code may continue processing a tool response even after a post-response block is reported.
Network inspection also doesn't work with agents that use certificate pinning or HTTP/3. Seeing an agent in the inventory does not mean runtime protection covers it, so check the support matrix.
Step 6: Purview DLP and the response process
For data protection, configure Purview Endpoint DLP for supported local agents. Start with a test policy, for example one that blocks a test sensitive-information pattern for the target agent. Wait for the policy to sync, confirm it is in effect on the endpoint, then test both pasting data directly into a prompt and having the agent read a test file. Create the policy once, centrally. Check receipt per device, but don't recreate the policy on each endpoint.
When something triggers (an MXC denial, a Defender alert or a Purview match), the guide's response process is:
- Detect the signal.
- Gather context: agent, user, device, process chain, tool, data type, destination and protection mode.
- Classify it as malicious, an expected workflow blocked by incomplete policy, a product gap, or a by-design limitation.
- Contain it with Defender response, Purview enforcement, identity controls, MXC policy or MDM exclusion.
- Fix and revalidate in a lower ring.
- Learn: update the inventory, policy rationale, support article and promotion criteria.
The classification step matters most. Without it, a SOC can spend hours on what turns out to be a missing folder grant.
Worked example: GitHub Copilot CLI
Microsoft walks through a governed Copilot CLI rollout:
- Identify the supported Windows build and the oldest Copilot CLI release you will allow.
- Enable the CLI and its managed settings for the GitHub enterprise.
- Create target and opt-out groups.
- Validate identity and security settings.
- Assign the CLI configuration and MXC policy to a small test group.
- Run Learning Mode across repository access, shell commands, tool calls, authentication, network calls, updates and failures.
- Turn the legitimate activity into a least-privilege policy and investigate anything unexpected.
- Expand through the rings, replacing temporary scripts with native endpoint management where possible.
Microsoft also says OpenAI has published its own guidance for Codex on Windows with MXC. eSecurityPlanet reported that Microsoft lists GitHub Copilot, OpenAI Codex, OpenClaw, Replit, LM Studio, NVIDIA OpenShell, and Unsloth AI among the agents and frameworks already supporting MXC. Treat the Copilot walkthrough as an illustration. Not every agent supports MXC, Defender runtime protection or Purview with the same controls.
Analysis: a solid framework, but read the preview labels
The framework is good. Observe, then enforce, roll out in rings, and test rollback before you need it: these are proven patterns applied to a new kind of workload. The split of responsibilities is also correct. Microsoft is explicit that MXC is one layer among several and doesn't overstate what containment can do.
Three cautions:
- Licensing adds up. The full design includes Defender for Endpoint Plan 2, Intune, Purview, Agent 365, and E7 or Agent 365 for risk scoring. This is an all-Microsoft stack, and the guide is written by the vendor that sells it. Organizations using other MDM or EDR products can use the method, but the integrations are Microsoft's.
- Preview means preview. Intune MXC policy is in private preview and Defender runtime protection is in preview with known issues. Plan pilots, not production timelines.
- Coverage varies. Nearly every capability comes with "supported agents" or "where supported." The real work is building the support matrix for your own environment.
Bottom line: Start the inventory now. Defender discovery runs automatically on eligible, onboarded devices. Run Learning Mode in a lab, keep Defender in Audit on a small group, and test rollback. Then, when Intune MXC policy reaches your tenant, you'll have the evidence to write a sensible policy instead of guessing.
References
- Enterprise Deployment of Governed Agents on Windows Windows Developer Blog · 2026-11-17T18:15:01+00:00
- Microsoft Execution Containers (MXC) GA: AI agent sandbox techaiwire.com
- Microsoft Execution Containers GA: AI Agent Security esecurityplanet.com