Systematic Analysis of MCP Security
Capture Summary
Paper introducing MCPLIB, a taxonomy and framework for MCP attacks. Search result reports 31 attack methods across direct tool injection, indirect tool injection, malicious user attacks, and LLM-inherent attacks.
Relevance
- Foundational MCP attack taxonomy.
- Useful for extending agent security topic pages from generic tools to MCP-specific surfaces.
- Candidate for evidence on context sharing, file-based attacks, and tool-description trust failures.
Collection Notes
Collected as current MCP security research.