mcp-builder
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The evaluation harness processes external test data which could contain prompt injection attempts. Ingestion points: The XML evaluation file parsed in
scripts/evaluation.py. Boundary markers: The system prompt uses XML tags to structure outputs. Capability inventory: The script can list and call tools on connected MCP servers and communicate with the Anthropic API. Sanitization: No explicit sanitization of input question text is performed. - [EXTERNAL_DOWNLOADS]: The skill fetches documentation and SDK information from the official Model Context Protocol GitHub repositories and the
modelcontextprotocol.iodomain. These are official resources for the protocol. - [COMMAND_EXECUTION]: The evaluation script and connection module facilitate the execution of local server commands via the stdio transport to enable testing of locally developed servers.
Audit Metadata