Securing Local AI Agents After Ignite 2026: Claude Code, Copilot CLI and the MCP Firewall
Back in April I wrote about the Mythos breach being a non-human identity problem. This week at Ignite in San Francisco, the non-human identity conversation moved somewhere much closer to home: the laptops of your developers.
Coding agents like Claude Code, GitHub Copilot CLI and OpenAI Codex, plus desktop assistants and local model runtimes, are on endpoints right now, reading files, calling tools, connecting to MCP servers and running commands. They do all of that with the signed-in user’s access. As Microsoft’s Defender documentation puts it, these agents “can’t reliably separate trusted content from hidden instructions”. One poisoned README or web page and an agent with your developer’s permissions is doing someone else’s bidding.
The good news is that this was one of the most joined-up stories at Ignite. Defender, Entra, Intune and Global Secure Access each have a piece, and Agent 365 pulls the inventory together. The less good news is that the status of each piece varies, so let’s go through it carefully.
Step One: Find Them
You can’t secure what you can’t see, and in my experience most organisations have no idea how many local agents are running.
According to the Agent 365 Ignite blog, Defender and Entra Global Secure Access together now discover 60+ types of local AI agent and their configured MCP servers. Defender covers managed Windows and macOS devices and ties each agent to a device and a user (the support matrix lists exactly which agents). Global Secure Access adds network-level visibility of each agent’s destinations, which Microsoft calls agentic traffic acquisition.
Status matters here:
- Local agent discovery in Defender is generally available.
- Network-level discovery in Global Secure Access and shadow AI discovery in the Agent 365 registry are preview. The Book of News lists agentic traffic acquisition as GA, though, so see the caveats below.
Step Two: Give Them an Identity
This is the bit I find most interesting. With Entra Agent ID for discovered local agents (preview), supported discovered agents get their own Agent ID, starting with GitHub Copilot CLI and Claude Code (Entra Ignite blog, Agent 365 Ignite blog). That separates what the agent does from what the human does.
Once agents have identities, Conditional Access can start treating them differently. Conditional Access policies for assistive agents (preview) can distinguish a person from a local agent acting on their behalf. So you can block a high-risk agent while the employee keeps working. There’s also a new actions and destinations option to restrict sensitive operations during risky sessions.
In my view, this is the right model. Blocking the user because their agent misbehaved is a blunt instrument, and blunt instruments get turned off.
Step Three: Protect Them at Runtime
Microsoft announced Defender runtime protection for local AI agents as generally available on Windows in the Defender & Sentinel Ignite blog, though the docs haven’t fully caught up (see the caveats below). Defender inspects critical points in the agent loop and can block malicious instructions before the agent acts.
There are two inspection methods, per the Ignite 2026 Book of News and the docs:
- Agent-native inspection uses vendor-supported event hooks. Announced as GA for Claude Code, GitHub Copilot CLI, the GitHub Copilot app, Claude Desktop (Code), and the VS Code GitHub Copilot and Claude Code extensions.
- Network-based inspection watches agent-to-LLM traffic. Announced as GA for OpenClaw, Ollama Desktop, ChatGPT Classic and ChatGPT Desktop (Chat).
You configure it in the Defender portal or Intune with Audit or Block mode. Every detection raises an alert with agent and session context. In Block mode, the user is also notified in the agent UI and through a Windows toast. Per the Book of News, it’s for Defender for Endpoint P2 customers.
Microsoft also announced new detections for attacks on the identities behind agents, in preview according to the Learn article. One example is device-code phishing that compromises a developer’s GitHub account in order to misuse GitHub Copilot CLI. Defender can automatically disrupt it by suspending the compromised account. Read the prerequisites before you count on it, though. Detection needs Microsoft 365 E5, Defender for Office 365 with Safe Links click tracking, and the Defender for Cloud Apps GitHub connector, and today both detection and disruption only work for GitHub Enterprise Managed Users (with disruption needing them federated with Entra ID).
Step Four: Contain Them on the Device
Intune local agent governance is in preview. Per the Book of News, Intune will be able to use Agent 365 discovery to mark devices running unapproved agents as non-compliant. Entra Conditional Access then blocks those devices from company data. For approved agents, Intune policies define boundaries using Microsoft Execution Containers (MXC), Microsoft’s open-source, policy-driven sandbox for limiting what an agent’s workloads can read, write and connect to.
That’s a proper closed loop: discover, decide, enforce through compliance and Conditional Access. It’s the same pattern we’ve used for unmanaged devices for years, now applied to agents.
Step Five: Control What They Talk To
The MCP firewall in Global Secure Access (preview) gives you policy over agent-to-MCP traffic. You can block all MCP traffic, allow-list servers, allow or block individual tools, resources and prompts, and block outdated protocol versions or unencrypted HTTP. The configuration guide walks through it. In short, you create an MCP policy, link it to a security profile, and enforce it through Conditional Access.
Read the prerequisites and limitations before you get excited, though:
- You need an Entra Internet Access licence, the GSA client on Entra joined or hybrid joined devices, and TLS inspection switched on.
- It inspects MCP over streamable HTTP and Server-Sent Events only. MCP over stdio isn’t supported, and local MCP servers on the device aren’t visible to GSA.
That last point is important. In my experience, a lot of developer MCP setups run local servers over stdio, so the MCP firewall covers remote MCP servers, not everything an agent does locally. That’s where Defender runtime protection and Intune containment need to pick up the slack.
Where I’d Start on Monday
- Turn on discovery and look at the numbers. Local agent discovery in Defender is GA, so check the inventory. I’d bet it’s bigger than anyone expects. Look at the connected MCP servers, and at agents that auto-approve their own actions. Defender reports that setting in the agent details and in the
AgentsInfoadvanced hunting table. - Enable runtime protection in Audit mode on a pilot group of Windows developer devices. Microsoft itself recommends starting in Audit and validating detections before you switch to Block. Runtime protection is Windows-only for now, so Mac users get discovery only. Plan their coverage separately, and scope any customer service with that in mind.
- Agree an approved agents list with your engineering leads. You’ll need it before Intune compliance or MCP allow-lists make any sense.
- Pilot the MCP firewall in a lab with a default-block policy and an allow-list of trusted servers, so you understand the TLS inspection impact.
- Plan for identity. Even in preview, decide how Agent IDs and agent-aware Conditional Access will fit with your existing CA design.
Caveats and Conflicting Information
- Runtime protection status. The Defender blog and Book of News say GA. But the Learn article still has “(Preview)” in its title as of 18 Nov, and the support matrix heads its runtime section “Windows (preview)”. Microsoft’s blog also marks the GA with an asterisk. The known issues are worth reading too: during preview, Claude Code might keep processing a tool response even in Block mode when Defender reports a post-tool-response block. I’d treat GA as applying to the agents and modes listed, and check the support matrix.
- Agentic traffic acquisition in Global Secure Access. The Book of News lists it as generally available, but the Entra blog lists it as preview. I’m going with preview until Microsoft clarifies.
- Discovery numbers. The Agent 365 blog says Defender and GSA discover “more than 60” local agents, while the Entra and Security blogs say 60 agent types. It’s a moving number either way, so check the support matrix.
- Licensing. Runtime protection needs MDE P2. The MCP firewall needs Entra Internet Access. Per the Book of News, full Agent 365 governance and security capabilities need an Agent 365 or Microsoft 365 E7 licence. Map this out before you promise anyone a solution.
- Platform. Runtime protection is Windows-only for now, while discovery also covers macOS. Plenty of developers are on Macs.
The MSP View
For those of us delivering managed security, this is a new service line, not just a feature. Customers are adopting coding agents faster than their security teams can write policy. “We’ll find every agent in your estate, tell you which ones are risky, and put guardrails around the ones you keep” is a conversation I’d want to be having in the next quarter. Just be honest about what’s preview.
Local agents are the newest kind of non-human identity, and they’re running with your users’ access. Treat them that way.