mcp-builder
Pass
Audited by Gen Agent Trust Hub on Jul 2, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The
scripts/evaluation.pyscript possesses an indirect prompt injection surface as it processes external XML data for test cases. - Ingestion points: The script reads user-provided XML files and extracts question strings from
<qa_pair>elements. - Boundary markers: Absent. Questions are interpolated directly into the agent's prompt.
- Capability inventory: The evaluation loop can invoke any tools provided by a connected MCP server.
- Sanitization: Absent. This is expected in a testing context where the user provides the test inputs.
- [COMMAND_EXECUTION]: The utility scripts provide standard local command execution capabilities.
- Evidence: The scripts allow launching local MCP servers via subprocesses using the
stdiotransport. This is a core feature for testing and integrating local tools. - [EXTERNAL_DOWNLOADS]: The documentation references official protocol resources.
- Evidence: References to
modelcontextprotocol.ioand the official GitHub repositories are present and point to trusted domains.
Audit Metadata