Benchmark your AI security readiness for 2027 → [Take the survey]

Benchmark your AI security readiness for 2027 → [Take the survey]

Benchmark your AI security readiness for 2027 → [Take the survey]

//Question

What Additional Security Do Claude Code, Cursor and MCP Tools Need Beyond Microsoft E5?

Posted on 19th November, 2026

Harry

Harry

//Answer

Microsoft E5 now covers more of the coding-agent stack than it did a year ago. Defender for Endpoint discovers supported local AI agents, including Claude Code and Cursor, along with their configured MCP servers, on Windows and macOS. It also adds runtime protection for coding agents that expose hooks. What’s left to evaluate is coverage depth: which agents, which actions and which data paths each control actually reaches in your environment.

Security layer

Microsoft baseline

Additional control to evaluate

Agent discovery

Defender discovers supported local AI agents and MCP configurations

Verify complete inventory across every IDE, CLI and local MCP server your developers use

Data protection

Purview provides DLP and AI data-governance controls, including blocking sensitive data headed to untrusted cloud services

Test sensitive-code and secret protection in each coding-agent workflow, including prompts and tool arguments

Agent permissions

Microsoft and vendor-native controls provide identity and policy capabilities

Enforce agent-specific limits on shell commands, file access and tool calls

MCP security

Defender provides MCP discovery and related visibility

Inspect MCP requests, tool outputs and risky actions before execution

Developer endpoints

Defender provides endpoint protection; its runtime protection documentation lists Claude Code, Codex CLI and GitHub Copilot CLI

Validate coverage for every agent you run, such as Cursor, plus local stdio MCP servers on Windows and macOS

The quickest way to find real gaps is to test, not compare feature lists:

  • Run a prohibited tool call in each coding agent and check whether it was blocked before execution or only logged.

  • Send a dummy secret through a prompt and through an MCP tool argument.

  • Confirm that inventory matches what’s actually installed.

Akto Atlas is designed to work alongside existing EDRs such as Defender. Akto can deploy its AI Endpoint Shield hook through Microsoft Defender Live Response, and uses Defender queries to find installed AI tools. It adds guardrails through the hook systems of Claude Code, Cursor, GitHub Copilot and OpenCode, inspecting prompts and tool calls before they run. MCP requests and responses get the same inspection through Akto MCP Proxy.

Treat E5 as the foundation, and fill gaps only where your own tests show them.

Comments