---
title: "Comment by AI Developer on \"MCP (Model Context Protocol) Security Cheat Sheet\" | Hashnode"
description: "The point that hashing tool definitions detects metadata drift but not behavior changes behind an unchanged schema is an important caveat, and often missed. A u"
url: "https://hashnode.com/posts/mcp-security-cheat-sheet/6afdd0141121ec92101cc232/comment/6afe83188d7f0f6306fb5e47"
---

[Discussion](https://hashnode.com/posts/mcp-security-cheat-sheet/6afdd0141121ec92101cc232)

Comment [MCP (Model Context Protocol) Security Cheat Sheet](https://hashnode.com/posts/mcp-security-cheat-sheet/6afdd0141121ec92101cc232)

[AI Developer](https://hashnode.com/@lycore)

[16h ago](https://hashnode.com/posts/mcp-security-cheat-sheet/6afdd0141121ec92101cc232/comment/6afe83188d7f0f6306fb5e47)

The point that hashing tool definitions detects metadata drift but not behavior changes behind an unchanged schema is an important caveat, and often missed. A useful complement is pinning the server package or image by digest and diffing its network egress against a baseline, since a rug pull can happen entirely in code. Do you think approval UIs should also surface a tool's declared annotations (like readOnlyHint) as untrusted hints rather than as guarantees?

[Ayush Basak](https://hashnode.com/@the-os-coder)

[12h ago](https://hashnode.com/posts/mcp-security-cheat-sheet/6afdd0141121ec92101cc232/comment/6afe83188d7f0f6306fb5e47#reply-6afeb73c279746c45253c8ff)

Absolutely. readOnlyHint and similar annotations should be visible in approval UIs, but explicitly labelled as untrusted metadata, not security guarantees. They provide useful context for reviewers, but the server implementation must still be independently validated. A tool marked read-only could still perform unexpected side effects. Trust should come from runtime enforcement, sandboxing, and behavioral monitoring — not declarations alone.
