mcp-builder
Pass
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill instructions and documentation references fetch content from official and well-known sources, including the Model Context Protocol website (
modelcontextprotocol.io) and its official GitHub repository. These fetches are for protocol specifications, README documentation, and SDK guidelines, which are necessary for the skill's primary function. - [COMMAND_EXECUTION]: The evaluation harness provided in
scripts/evaluation.pyandscripts/connections.pyallows for the execution of local MCP servers via a subprocess when using thestdiotransport. This is the standard operational model for testing MCP servers and is performed locally under the developer's control. - [INDIRECT_PROMPT_INJECTION]: The
scripts/evaluation.pytool possesses an indirect prompt injection attack surface as it processes external data. - Ingestion points: The script reads evaluation tasks (questions and ground truth answers) from an XML file provided by the user via the CLI.
- Boundary markers: The
EVALUATION_PROMPTsystem prompt establishes clear structural boundaries using XML tags (<summary>,<feedback>,<response>) to separate instructions from the data being processed. - Capability inventory: The script can list and call tools on the target MCP server using the official
mcpclient library. - Sanitization: Output processing uses regular expressions to extract content from tags, ensuring logical separation between the AI's reasoning and the final response.
Audit Metadata