Version 3 is out now!
A practical, community-driven checklist for pentesting Model Context Protocol (MCP) servers. This guide covers local and remote MCP server risks, traffic analysis, tool-call behaviors, context boundaries, authorization flows, and unsafe code paths.
Originally created for the OWASP Bay Area talk on Pentesting MCP Servers (Oct 2025), this checklist is designed for practitioners performing assessments on MCP-based tools, agents, and integrations.
MCP servers are becoming the new execution layer for AI agents. This means they expose:
- File system access
- Tool execution
- Remote APIs
- STDIO and HTTP bridges
- Autonomous actions initiated by LLMs
Because of this, MCP servers introduce a wide attack surface that security testers need structured guidance for. This checklist helps you perform systematic and repeatable assessments.
- Traffic Analysis — proxy inspection of STDIO/HTTP, context injection, TLS enforcement
- Authentication & Authorization — auth bypass, OAuth flows, IDOR, privilege escalation
- Local MCP Server File and Code Review — embedded secrets, dangerous functions, dependency audits
- MCP Tool Behavior and Functionality — tool boundary validation, chaining, local RCE
- Tool Security — Input Validation — command injection, path traversal, SSRF, SQLi, SSTI
- Tool Security — Output & Schema Validation — schema mismatches, sensitive data leakage, prompt injection via output
- Tool Injection — prompt injection via tool names/descriptions, tool shadowing, name collisions
- File System & Network Access — path traversal, scope enforcement, DNS rebinding
- Context Isolation — cross-user leakage, namespace separation, session persistence
- Secret & Credential Handling — hardcoded secrets, log exposure, token caching
- Logging & Monitoring — log injection, rate limiting, access controls
- Race Conditions & Concurrency — TOCTOU, parallel invocation, resource exhaustion
- Advanced Attacks — context pollution, confused deputy, prototype pollution, GraphQL injection
- Use the CHECKLIST.md for field assessments
- Fork and adapt it for your team
- Submit PRs with improvements
- Open issues for new MCP attack patterns
We welcome:
- New checklist items
- Additional MCP server categories
- Tooling contributions
- Red-team test cases
- Sanitized findings
This project is licensed under CC BY 4.0. You may remix, adapt, and build upon this checklist for any purpose, even commercially, as long as you provide attribution.
Appsecco