Microsoft Execution Containers: Agent Security, Explained

Microsoft Execution Containers (MXC) are now generally available, giving Windows an operating-system-enforced containment layer for AI agents. Instead of trusting an agent to stay inside its lane, MXC applies least-privilege controls over files, networks, processes and user interaction at runtime - enforced by the OS, outside the agent's own control.
- Generally available on Windows — Microsoft's November 2026 security update states that Microsoft Execution Containers are now generally available, providing an operating system-enforced containment layer for AI agents with least-privilege controls across files, networks, processes and user interactions.
- Task-scoped permissions — Dynamic Permissions gives an agent its own permissions scoped to its current task and only for as long as the task requires, while Dynamic Agent Risk assesses whether that authority still matches behaviour and context.
- Four containment backends — MXC offers a process container on Windows, macOS and Linux, a Windows-only session container for long-running agents, a Windows-only WSL container for Linux-first toolchains, and an experimental MicroVM for hardware-backed isolation on Windows 11 and Linux.
- Agents already wired up — Microsoft lists GitHub Copilot, OpenClaw, OpenAI Codex, Replit, LM Studio and Unsloth AI as already supporting MXC, with Anthropic Claude Code, Box, Egnyte, Manus, Perplexity, Raycast and others named as coming.
What "generally available" actually means
Related guide: Claude Code Mods: How Anthropic's TypeScript Agent Hooks Actually Work
Microsoft's November 2026 security update is the authoritative status line: MXC is now generally available and gives organisations "an operating system-enforced containment layer for AI agents, designed to limit what an agent can access". Rather than relying on the agent to behave within intended boundaries, MXC applies least-privilege controls across files, networks, processes and user interactions, with additional isolation options for higher-risk scenarios. Organisations can tune and validate policies before enforcing them broadly, and upcoming integration with Microsoft Intune, Microsoft Entra and Microsoft Agent 365 is intended to extend governance and attribution across the enterprise.
The Windows developer blog, published October 7, 2026, describes the same capability from the developer's side. Developers and IT administrators define the resources a workload needs - files and network destinations - and MXC enforces the resulting boundary using the appropriate container. The policy "remains outside the agent workload's control, so the agent or generated code cannot grant itself additional access". The implementation is a unified JSON configuration schema plus multi-language SDKs, with MXC mapping the requested controls onto the selected backend on Windows, macOS or Linux. The SDK, schema and samples are published in Microsoft's MXC repository on GitHub.
Why containment matters more than prompts
Related guide: IBM and Microsoft Launch Imagine Enterprise Agent Platform
Microsoft's own worked example is worth reading, because it describes a failure that is not a bug. Consider a coding agent asked to update a website. It needs read and write access to the repository and to the build and test tools. It may need to read production server configuration to understand deployment. It should not be able to modify that configuration.
The point is that an agent might decide that changing the server configuration is the fastest way to complete the task. "The action can be reasonable from the agent's perspective and still exceed the authority the developer intended to grant." A prompt-level instruction - "do not touch the production config" - cannot close that gap, because the instruction and the enforcement share the same authority. MXC closes it: the agent can read and write the repository and read the server configuration, but if it attempts the modification the containment environment is designed to prevent the operation regardless of what the model, generated code, plugin or tool decides to do.
The four isolation levels
Related guide: Gemini 4 Hydrogen: Google's faster agent model explained
MXC separates workload requirements from platform-specific containment detail, and offers a spectrum rather than a single mode. Choosing the wrong one is usually over-isolating; choosing too little is the usual failure.
- Process container - available on Windows 11, macOS and Linux. The lightweight option for responsive workloads, including model-generated code and tool execution. It uses the platform's native process sandbox: AppContainer on Windows, Seatbelt on macOS and Bubblewrap on Linux.
- Session container - Windows 11 only, and the only backend of its kind. It runs the agent on the user's device in a separate, OS-isolated session with its own local agent identity and isolated desktop, clipboard, UI and input boundaries. Microsoft positions it for long-running agents and automation that need a desktop or stronger separation from the interactive user.
- WSL container - Windows 11 only, for Linux-first agent toolchains and workloads that depend on the Linux package and development ecosystem.
- MicroVM - Windows 11 and Linux, marked experimental. It provides hardware-enforced isolation with full Linux workload compatibility, for higher-risk workloads.
Windows 365 support for MXC is also generally available, so agents can run on Cloud PCs with the same isolation model rather than only on a developer's laptop.
Which agents already support MXC
Related guide: Self-Hosted Claude Managed Agents: What It Changes
The integration list is the most immediately useful part for anyone choosing a coding agent today. Microsoft names GitHub Copilot, OpenClaw, OpenAI Codex, Replit, LM Studio and Unsloth AI as already supporting MXC. On the "will release support" list are Anthropic Claude Code, Box, Egnyte, Heidi Health, Hermes Agent by Nous Research, Manus, Perplexity, Raycast and Simular.
For coding workflows, Microsoft highlights GitHub Copilot, Codex and Replit as the demanding cases: the agent needs the repository, development tools and commands, but should not automatically gain access to unrelated files or network destinations. NVIDIA has separately integrated OpenShell into MXC, adding policy controls for access to files and inference services, advanced network controls, credential management and, for enterprises, OCSF auditing.
One honest caveat: "will release support" is not the same as shipping today, and Microsoft's own November roundup notes that Intune management for MXC process containers on Windows 11 is still listed as coming, without a date. Treat fleet-wide policy deployment as something to verify rather than assume.
The wider stack: identity, risk and discovery
Related guide: NVIDIA RTX Spark desktops: local AI on Windows
Containment is one layer in a broader Microsoft push that is also landing this month and being showcased at Microsoft Ignite, which runs November 17-20 in San Francisco and online. Three companion pieces are worth tracking.
Dynamic Permissions and Dynamic Agent Risk give an agent its own permissions, scoped to its current task and only for as long as the task requires. Dynamic Agent Risk then assesses whether that authority still matches the agent's behaviour and context, and Dynamic Conditional Access can respond proportionally - narrowing permissions, restricting access to resources, or requiring human approval for higher-impact actions.
Defender now discovers local agents on Windows and macOS. Microsoft's November update says the capability is generally available and that Defender "discovers more than 50 supported local AI agents by observing their activity", bringing the agent, device, account and configured MCP context together so security teams can understand local agent activity. Discovery is the precondition for everything else: you cannot apply least privilege to an agent you cannot find.
Agent 365 is the observability and governance layer above all of this. Microsoft Learn describes it as the observability platform for AI, built on four pillars - management, governance, security and business value - with an agent registry in the Microsoft 365 admin center, approval workflows for publishing agents, policy templates, Purview activity records for audits, and Copilot Credit cost management broken down by user, group, agent and service. It is generally available for the Commercial segment on a per-user basis. Microsoft's Agent 365 blog for Ignite frames the year ahead as moving the product "from an agentic control plane into an AI observability platform".
What to do with this
Related guide: Claude Sheets and Whiteboards: What Teams Get
If you run agents on Windows today, the practical sequence is short. Start by identifying what your agents actually touch - a coding agent and a research agent have very different resource footprints, and a container policy is only as good as that inventory. Then pick the lightest backend that still covers the blast radius: process container for most tool-call workloads, session container for long-running automation that needs a desktop, MicroVM only when the workload genuinely runs untrusted code.
If you are on macOS or Linux, note the asymmetry: the process container is available on all three platforms, but the session container and WSL container are Windows-only and MicroVM is experimental. For cross-platform teams, the uniform JSON schema and SDKs mean you declare requirements once and let MXC choose what the host can enforce.
The single thing to avoid is treating "the agent was told not to" as a security control. MXC exists precisely because that is not a control.
Frequently asked questions
Generally available on Windows
Microsoft's November 2026 security update states that Microsoft Execution Containers are now generally available, providing an operating system-enforced containment layer for AI agents with least-privilege controls across files, networks, processes and user interactions.
Task-scoped permissions
Dynamic Permissions gives an agent its own permissions scoped to its current task and only for as long as the task requires, while Dynamic Agent Risk assesses whether that authority still matches behaviour and context.
Four containment backends
MXC offers a process container on Windows, macOS and Linux, a Windows-only session container for long-running agents, a Windows-only WSL container for Linux-first toolchains, and an experimental MicroVM for hardware-backed isolation on Windows 11 and Linux.
Agents already wired up
Microsoft lists GitHub Copilot, OpenClaw, OpenAI Codex, Replit, LM Studio and Unsloth AI as already supporting MXC, with Anthropic Claude Code, Box, Egnyte, Manus, Perplexity, Raycast and others named as coming.
Prices and benchmark figures are standard published rates as of November 16, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.