CrowdStrike taxonomy: three attack classes targeting MCP server tool descriptions

voidrane · reddit · 2026-09-18

CrowdStrike's research team documented three attack classes against MCP tool descriptions: tool poisoning (hiding exfiltration instructions in metadata), tool shadowing (one tool's description contaminating how an agent uses another), and rugpull attacks (servers changing tool behavior post-integration via dynamic updates). Key insight: these vulnerabilities live in the relationship between tool descriptions and LLM inference, not in code, so static analysis misses them entirely. Mitigations include description pinning, context isolation, and OS-level least privilege.

Original post →

More from coding & agent

coding & agent channel →