mcp-pipeline-builder

Warn

Audited by Gen Agent Trust Hub on Mar 23, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through its ingestion of external repository content. Malicious instructions embedded in a target repository's documentation or source code could influence the analysis and design phases, potentially leading to the generation of malicious code. \n
  • Ingestion points: Phase 1 repository analysis in SKILL.md. \n
  • Boundary markers: None identified during repository ingestion; the skill relies on user review. \n
  • Capability inventory: Shell access (Bash), package installation (npm/pip), subprocess execution (Phase 5), and config modification (Phase 6). \n
  • Sanitization: A mandatory human review gate in Phase 2 requires user approval before code generation.\n- [EXTERNAL_DOWNLOADS]: The skill downloads content from user-provided repository URLs during Phase 1 using git clone. In Phase 4, it performs npm install, which fetches third-party dependencies from the npm registry based on generated package.json files.\n- [REMOTE_CODE_EXECUTION]: Phase 5 involves launching the newly generated and compiled MCP server as a subprocess for evaluation. This constitutes execution of code generated from potentially untrusted external inputs.\n- [COMMAND_EXECUTION]: The skill uses the Bash tool to execute several high-impact commands, including repository cloning, package installation (npm install), and build processes (npm run build, python -m py_compile).\n- [DATA_EXFILTRATION]: Phase 6 modifies the user's global Claude configuration file (~/.claude.json) or local settings to register the new server. While this is the intended functionality, it involves programmatic modification of sensitive system configuration.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 23, 2026, 04:37 PM
Security Audit — agent-trust-hub — mcp-pipeline-builder