cd /news/ai-agents/microsoft-defender-now-guards-claude… · home › topics › ai-agents › article
[ARTICLE · art-172456] src=digitalapplied.com ↗ pub= topic=ai-agents verified=true sentiment=· neutral

Microsoft Defender Now Guards Claude Code and Copilot CLI

At Ignite on November 17, Microsoft announced general availability for Defender runtime protection of local AI agents on Windows, covering activity exposed by supported agents including Claude Code and GitHub Copilot CLI and giving security operations centers agent and session context. As of November 19, Microsoft's support matrix and setup guide still label Windows runtime protection as preview, and the operational overview and configuration guide carry preview labels and known issues, so Microsoft's announced status alone cannot resolve a documented enforcement limitation. Local agent discovery is already generally available on supported Windows and macOS devices, and Microsoft says Defender and Entra Global Secure Access together discover more than 60 local agents.

key takeaway

Defender's Windows runtime protection for local AI agents is announced as generally available but still documented as preview, and because Audit mode can show a misleading block title and Claude Code has a post-tool-response limitation in Block mode, analysts must check the endpoint timeline and the

read11 min views3 publishedNov 19, 2026
Microsoft Defender Now Guards Claude Code and Copilot CLI
Image: Digitalapplied (auto-discovered)

A security analyst opens a Defender alert that says a suspicious AI prompt was blocked on a developer’s laptop. The developer was using Claude Code. Before closing the incident, the analyst needs two facts the headline cannot supply. Was that machine in Audit or Block mode? Did the detection occur before the tool ran or after it returned its result? Microsoft documents cases where the answer to either question changes what the team can conclude.

At Ignite on November 17, Microsoft announced general availability for runtime protection of local AI agents on Windows. Defender can inspect activity exposed by supported agents, including Claude Code and GitHub Copilot CLI, and provide the security operations center with agent and session context. As of November 19, its support matrix and setup guide still label Windows runtime protection as preview. Teams can evaluate the control now while checking that their endpoint, agent interface and required assurance level match what Microsoft documents. Microsoft’s Ignite announcement · Defender’s agent support matrix

  1. 01Match the application and mode to a Windows runtime entryNative event inspection covers Claude Code, Copilot CLI and specific desktop or editor integrations. A separate group uses network inspection. The Windows and macOS discovery inventory covers more products than either group.
  2. 02Audit allows the action and still alerts the security teamMicrosoft recommends one to two weeks of reviewing alerts on a small active group. Move selected device groups to Block after validating that the detections are accurate and actionable.
  3. 03A block alert needs an endpoint timelineMicrosoft documents a misleading block title in Audit mode and a separate Claude Code post-tool-response limitation in Block mode. An alert title by itself cannot establish prevention.
  4. 04Decide whether to collect prompt snippetsFlagged snippets may include unredacted secrets or conversations. Evidence collection has a separate device setting and update-dependent defaults. Its setting does not switch runtime protection on.

01 — Release statusWhat reached general availability? #

There are two release records to keep. The Ignite article marks Windows runtime protection as generally available with an asterisk, without explaining that mark. The operational overview and configuration guide still carry preview labels and known issues. The matrix also labels its Windows runtime tables preview. If your rollout requires a GA support commitment, keep the announcement with the applicable deployment documentation and confirm the support status for your tenant with Microsoft. The announced status alone cannot resolve a documented enforcement limitation. Microsoft Learn’s runtime overview and known issues

Local agent discovery is already generally available on supported Windows and macOS devices, with exceptions identified in the matrix. Discovery inventories local agents and, for marked entries, their configured Model Context Protocol (MCP) servers. Microsoft says Defender and Entra Global Secure Access together discover more than 60 local agents. That total refers to discovery across the two products. An inventory record identifies software or activity on a device. It cannot tell you an action was inspected or stopped. Microsoft’s Agent 365 discovery announcement · Discovery prerequisites and inventory behavior

Two other Ignite announcements concern different risks. Assessment of agents built on third-party cloud platforms, including Amazon Bedrock, Google Vertex AI and Anthropic Claude, is in preview via Agent 365 connected platforms. Microsoft also described detection and disruption for GitHub account compromise involving device-code phishing. Neither an inventory entry for a cloud agent nor a response against a compromised account confirms runtime coverage for the local coding tool on a laptop.

02 — Support boundariesWhich agent is running, and where? #

Start with the runtime section of the support matrix. Microsoft lists combinations by application, operating system, mode and inspection method. Claude Desktop’s Code tab and ChatGPT Desktop’s Chat mode belong to different rows and use different methods. The product logo on a laptop or a general reference to a CLI in an overview does not establish a supported combination. Windows runtime support by application and method

Claude Code and GitHub Copilot CLI

The matrix also lists the GitHub Copilot app, Claude Desktop in Code mode, the VS Code GitHub Copilot extension and the VS Code Claude Code extension. Defender uses the event interfaces the vendor exposes. The matrix does not specify a minimum agent version for those six entries.

ChatGPT Desktop in Chat mode

The minimum tested ChatGPT Desktop build is 26.810.52044. OpenClaw, Ollama Desktop and ChatGPT Classic also have network-inspection rows. Microsoft limits support to covered inference-service connections and excludes certificate-pinned and HTTP/3 traffic.

Discovery has a wider reach

The matrix gives macOS inventory coverage, with specific VS Code extensions in preview. The Windows Subsystem for Linux has preview discovery for Claude Code, Codex CLI and GitHub Copilot CLI. The runtime tables contain no macOS, WSL or native Linux entries.

Codex CLI illustrates why the distinction matters. Microsoft’s overview mentions it as an agent with a vendor event interface, and its Windows discovery row includes MCP configuration. The November 19 Windows runtime table does not list it. If your engineering team uses Codex CLI, record its discovery status separately and ask for product-specific confirmation before treating a Windows runtime policy as coverage. Which coding agents expose tool-call hooks?

03 — ReadinessCheck the licence and the endpoint #

The runtime setup guide lists Defender for Endpoint Plan 2, Microsoft 365 E5, Microsoft Agent 365 or Microsoft 365 E7. It requires a supported Windows device onboarded to Defender for Endpoint, Microsoft Defender Antivirus in active mode with real-time protection on, current platform, engine and security intelligence updates, and a supported installed agent. Its local setup procedure specifies antivirus security intelligence signature version 1.451.224.0 or later. An employee’s Copilot or Claude subscription does not supply these endpoint prerequisites. Microsoft’s runtime configuration and licensing guidance

The separate discovery feature requires at least Defender for Endpoint Plan 2 in the commercial cloud. The minimum Windows antivirus platform for that inventory is 4.18.26090.9, and the minimum macOS Defender client is 101.26072.0017. Those inventory version numbers do not certify a runtime deployment. Plan 2 includes the inventory and advanced hunting. Discovery risk indicators and recommendations require Microsoft 365 E7 or Agent 365 together with Defender for Endpoint Plan 2.

A small device trial can be configured locally. For a device group, Microsoft documents Intune or the Defender portal. Intune requires its own entitlement and the relevant device-configuration role. The Defender portal requires endpoint-security policy permissions. The analyst who reviews alerts can hold a security-read role without being the administrator who assigns the policy. Give each owner the agent name, device group, inspection method and intended mode so they can check the same rollout.

04 — EnforcementWhen to move from Audit to Block #

Defender exposes separate settings for agent-native inspection and network inspection. Each accepts Disabled, Audit or Block. Disabled turns off that inspection method. Audit records detected activity and sends an informational alert while permitting the action to continue. Block attempts to stop the detected activity where the inspection method supports it. The user can see an in-agent message and, if enabled, a Windows notification. Block alerts are assigned severity according to the assessed risk.

Microsoft recommends beginning in Audit on a small group that actively uses supported agents, reviewing alerts for one to two weeks, expanding Audit, then selecting groups for Block. Choose a group that represents ordinary work. A team that only reads code will tell you little about a group that edits files, runs builds and uses internal connectors. Have someone from developer tooling review legitimate work that produced an alert and have security review the same session for actual risk. Record the policy effective on the device. A group assignment alone does not show that a client received it.

In deciding whether to promote a group, ask whether users can complete their normal tasks, which alerts still need classification and who will answer a developer when Block interrupts a release. Microsoft provides a benign demonstration for validating the control on a test device. That can establish the alerting and notification path for the tested combination. It cannot measure a detection rate for the organization’s repositories or tools.

05 — Worked incidentHow to read a block alert #

Take a fictional ten-laptop pilot. Six developers use Claude Code in a Windows terminal, two use Copilot CLI on Windows, and two use Claude Code on Macs. The inventory can show ten installations. Only the eight Windows installations map to this runtime policy, assuming the required prerequisites and native inspection setting. List eight covered pilot devices and two inventory-only devices. If a Windows developer later moves the same work into WSL, revise the classification for that session.

On Tuesday the SOC receives a block-labelled alert for one Windows Claude Code session. The applied policy was Audit at the time. Microsoft documents a reporting issue in which the portal can label an Audit detection as blocked. Record the event as detected with the action allowed by policy. Check the affected account, agent and session, the process tree, device timeline and any resulting file or service changes before deciding what happened. The monitored developer can often identify the legitimate task and its expected outputs. Microsoft’s documented alert and enforcement limitations

On Thursday the same group has moved to Block. A Claude Code alert now concerns a response returned by a tool. Microsoft warns that Claude Code can continue processing that response despite a reported block. The tool has already run when this inspection occurs. Confirm what it did using the relevant repository or service’s records and check for later actions in the session. If the evidence shows only that the detector fired, leave prevention unconfirmed in the case record. That is a different finding from the Audit reporting issue.

Record the device and agent surface, effective inspection mode, event timing, observed tool outcome and independent evidence of the resulting state. For the Thursday example, an alert about a tool result and an unchanged repository establish two facts. They still leave open whether the tool touched an external system it was permitted to use.

06 — Evidence accessWho can see the flagged text? #

Prompt evidence collection is an independent device setting. When enabled along with Audit or Block, Defender can include portions of user prompts or agent responses that it identifies as suspicious and relevant to the alert. Microsoft says sensitive information in those snippets is unredacted. A support engineer investigating a detection could therefore see part of a developer’s conversation or an internal value. Decide who can read the evidence and whether it belongs in a ticket, and apply the organization’s existing security-data retention and access rules. Defender alert evidence follows Microsoft Defender XDR retention policy. Microsoft’s prompt evidence and privacy notes

The collection control requires the October 2026 antivirus platform and engine families 4.18.26100.x and 1.1.26100.x or later. Microsoft says a platform release will turn collection on by default, while devices that have yet to receive the change default to off. An explicit policy takes precedence. Disabling collection leaves runtime inspection working, and content associated with a blocked activity can still appear in its alert. A blanket statement that disabling the option removes all prompt material from the SOC would mislead both administrators and employees.

07 — Agent permissionsWhat should still require a developer’s approval? #

Agent permissions define which tools a user has authorized and which need approval. Endpoint inspection examines supported events or connections for suspicious activity. A permitted command can still be inappropriate in a particular task. An ordinary-looking request can also exceed someone’s authority without resembling a prompt-injection attempt. For a code-review assistant, permission to read the working tree does not inherently require permission to publish a release or edit repository settings.

Claude Code supports tool-specific allow, ask and deny rules as well as organization-managed policy. Copilot CLI provides trusted directories and permissions for tools, paths and URLs. A developer can trust the checkout for a session while still requiring approval for a proposed action outside the review. Give the pilot a clear written description of which approvals belong to the human and which operations the account is entitled to perform. Revisit those native permissions if engineers add a new connector with broader access. Anthropic’s permission model · GitHub’s CLI permissions guide · How agent permission defaults change a rollout

Use the smallest account and connector permissions that still let developers finish the intended job. The security team can then evaluate Defender against the operations the agent actually needs. When an organization expands its coding-agent deployment, two related decisions are how administrators manage editor configuration and how approval settings vary by tool. Governing AI coding in VS Code · Coding-agent permission defaults

Start with the devices you can verify

For supported Windows machines running Claude Code or GitHub Copilot CLI, Audit gives security teams evidence about real developer sessions before they enable enforcement. Move a group to Block when its policy, user experience, evidence access and incident response have been tested. Keep macOS, WSL and unlisted application modes visible as separate inventory groups. Our AI transformation services cover the permission design and deployment decisions behind an enterprise agent rollout.

── more in #ai-agents 4 stories · sorted by recency
── more on @microsoft 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/microsoft-defender-n…] indexed:0 read:11min 2026-11-19 · —