The MCP project's security guidance covers confused deputy proxies, token passthrough, SSRF during OAuth discovery, state handle hijacking, local server compromise, malicious authorization URLs, and scope minimization. Independent research adds tool poisoning, where instructions hide in tool descriptions, and cross-server attacks. You can use both to review any server before approval.
Read in this order
- OfficialSecurity Best Practices (Model Context Protocol)Start here: the official attack list with a mitigation for each.
- ArticleMCP Security Notification: Tool Poisoning Attacks (Invariant Labs)The original tool poisoning demonstration, with concrete malicious descriptions.
- OfficialA Practical Guide for Securely Using Third-Party MCP Servers 1.0 (OWASP GenAI Security Project)A checklist for vetting and running third-party MCP servers.