Survey finds 59 MCP tools that read untrusted input and act

What the research found

A static survey of 1,384 tools in 28 popular Python MCP servers finds 59 tools that both ingest untrusted content and take high-impact actions, with benign imperative tool descriptions indistinguishable from injections.

Claims and how well the evidence supports them

1. Survey covered 1,384 tools across 28 Python MCP servers 99%
2. 59 tools were both untrusted-content sources and high-impact sinks 99%
3. 9 pattern-flagged imperative descriptions were all benign on hand review 98%
4. Snapshot taken 2026-10-20 and no servers were executed 99%

Evidence

• github.com https://github.com/agentsploit/agentsploit/blob/main/docs/mcp-tool-survey.md

Actionability rated medium.