skills/hkuds/deepcode/mcp-builder/Gen Agent Trust Hub

mcp-builder

Pass

Audited by Gen Agent Trust Hub on Nov 7, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The evaluation script (scripts/evaluation.py) reads user-provided XML files containing test questions which are then interpolated into a system prompt for the AI agent. This creates a surface for indirect prompt injection if the evaluation files are sourced from untrusted parties. * Ingestion point: scripts/evaluation.py parses eval_file. * Boundary markers: Uses an EVALUATION_PROMPT system message to wrap the task. * Capability inventory: Can call any tool exposed by the MCP server under test and communicates with the Anthropic API. * Sanitization: No explicit sanitization or filtering of the question text is performed before it is sent to the LLM.
  • [COMMAND_EXECUTION]: The connection utility (scripts/connections.py) and evaluation script utilize the stdio_client to launch and communicate with local MCP servers. This involves executing arbitrary commands and arguments provided by the user via command-line interface.
  • [EXTERNAL_DOWNLOADS]: The documentation and skill instructions reference and fetch content from the official Model Context Protocol website and its public GitHub repositories. These downloads are limited to documentation files (READMEs and specifications) and utilize well-known, authoritative sources for the technology.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 7, 2026, 06:31 AM