Home › Security
MCP Security
This cluster paraphrases the Model Context Protocol specification dated 2025-06-18. It is not a penetration test of the servers in the directory, and it does not add incidents that the specification does not describe. The primary document is Security Best Practices. Authorization rules are in Authorization. Tool confirmation and annotation rules are in Tools.
- MCP Security Checklist: 12 Things to Check. Twelve controls copied in substance from the security best practices, each named back to its section.
- Prompt Injection via MCP Tools. The specification's session-hijack prompt injection case, plus the tools chapter's human-in-the-loop rule.
- Least-Privilege Tokens for MCP. Scope minimization and the rule that a token must be issued for that MCP server.
The tester and a Verified on MCPwell line only tell you what initialize and tools/list returned. They do not tell you that a server follows this checklist. Transport rules for Origin checks and localhost binding are summarized on What is MCP?, from the transport specification.
FAQ
Does MCPwell test these controls on each server?
No. The fact-sheet check sends initialize and, when that returns a result, tools/list. It does not test OAuth consent, token audience, session binding, or local-command consent.
Where do these points come from?
The 2025-06-18 security best practices at https://modelcontextprotocol.io/specification/2025-06-18/basic/security_best_practices, the authorization specification at https://modelcontextprotocol.io/specification/2025-06-18/basic/authorization, and the tools specification at https://modelcontextprotocol.io/specification/2025-06-18/server/tools.