A New Kind of Shadow IT
For twenty years, security teams fought shadow IT: apps and cloud services employees adopted without asking, quietly holding company data. The fix was always the same. Find them, then decide what to allow.
Local AI agents are the next wave, and they are harder. A coding assistant, a desktop copilot or an agent connected to company tools through Model Context Protocol (MCP) servers installs in minutes. Once running, it can read files, run commands, open network connections and act with the user's own permissions.
The difference with classic shadow IT is simple: an unapproved app stores data. An unapproved agent takes actions.
What Microsoft Shipped
Defender Finds Local AI Agents
Microsoft Defender now discovers more than 50 supported local AI agents on Windows and macOS by observing their activity. It also maps the MCP servers each agent is configured to use, and links together the agent, the device and the user account. For a security team, that is the first real inventory of which agents run where, and what they are plugged into.
Microsoft Execution Containers (MXC)
MXC is an operating system-enforced containment layer for AI agents on Windows, designed to limit what an agent can access. It applies least-privilege controls to files, network, processes and user interactions, with stronger isolation for higher-risk scenarios. Teams can tune and validate policies before enforcing them broadly. Integration with Intune, Entra and Agent 365 is announced as upcoming.
In the same announcement, Microsoft described where it wants to go next: dynamic permissions that give an agent access scoped to its current task, and only for as long as the task lasts; a dynamic risk score that checks whether an agent's authority still matches its behavior; and conditional access that can narrow permissions or require human approval for higher-impact actions. Microsoft did not state availability dates for these.
Why It Matters
An agent running with a user's permissions is, in practice, an insider that never sleeps. It can be wrong, and it can be manipulated. If a web page, an email or a document injects instructions into it (prompt injection), the attacker inherits everything the agent can touch: files, tokens, internal tools behind MCP servers.
- You cannot protect what you have not inventoried. Most organizations today cannot answer a simple question: which AI agents are running on our endpoints right now?
- MCP servers widen the blast radius. Each server an agent connects to is a new path to data or actions, often configured by the user, not by IT.
- Agents blur attribution. When an agent sends an email or deletes a file, the logs show the user. Tying actions to the agent is exactly what Microsoft's announced integrations aim at.
A Six-Step Plan for Security Teams
| Step | What to do |
|---|---|
| 1. Inventory | List which AI agents run on your endpoints and which MCP servers they connect to. Defender customers can start with the new discovery view; others can begin with software inventory, process telemetry and config files. |
| 2. Assign owners | Every approved agent needs a named human who answers for it, its permissions and its connections. |
| 3. Least privilege | No agent should hold standing admin rights or broad file system access. Scope access to the task, not to the user. |
| 4. Watch the network | Baseline where agents talk. A local agent opening outbound connections to unknown hosts deserves an alert. |
| 5. Human approval | Require a person to confirm high-impact actions: deleting data, sending email, moving money, changing configurations. |
| 6. Test before blocking | Run containment policies in audit mode first, measure what would break, then enforce. |
The Fine Print
The bigger point stands regardless of vendor: local agents are now endpoints with privileges, and they need the same lifecycle as any other privileged software.
On November 5, 2026, Microsoft made two features generally available: Defender discovery of more than 50 local AI agents and their MCP servers on Windows and macOS, and Microsoft Execution Containers (MXC), an OS-enforced layer that limits what AI agents can access on Windows. Together they treat local agents as a new form of shadow IT that must be discovered, owned and contained.
Frequently Asked Questions
My Take
The real news here is not a product launch. It is that the largest security vendor on Windows now treats local AI agents as endpoints to discover and contain. That is a signal for every security policy, whatever stack you run.
As a network security engineer, I would start with the network. Agents need to talk to models, tools and MCP servers, and that traffic is visible. Knowing which machines talk to which AI services is the cheapest inventory you can build this week, and it does not depend on any vendor.
Do you know how many AI agents are running in your company right now?
Sources:
This article is published for information purposes only, to help readers understand emerging AI security risks.
Related Articles: