secure-mcp-install

Warn

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
  • REMOTE_CODE_EXECUTION (MEDIUM): The skill implements a workflow that clones repositories from arbitrary URLs and executes their installation scripts (e.g., npm install, pip install -e .). This 'download then execute' pattern is a high-risk capability; however, because it is the primary intended purpose of the skill, the severity is reduced to MEDIUM.
  • EXTERNAL_DOWNLOADS (MEDIUM): The skill pulls source code from user-provided URLs (git clone <REPO_URL>) without restriction to trusted domains. This exposes the system to unverified external content.
  • COMMAND_EXECUTION (MEDIUM): The skill executes various system commands, including building external projects (npm run build) and modifying the agent's MCP configuration (claude mcp add-json). These operations could be subverted if the cloned repository contains malicious build hooks.
  • PROMPT_INJECTION (LOW): The skill is susceptible to indirect prompt injection because it is designed to ingest and analyze untrusted third-party source code. Attackers could place malicious instructions in code comments or metadata to deceive the agent during the audit process.
  • Ingestion points: Third-party repositories cloned to ~/.claude/mcp-audits/ (documented in SKILL.md).
  • Boundary markers: No specific delimiters are used to wrap or isolate the content of the files being audited.
  • Capability inventory: Subprocess execution for installation (npm, pip, python) and modification of the local configuration via the claude CLI.
  • Sanitization: The skill mitigates this risk through an automated scanner (audit-mcp-server.sh) and a detailed manual audit checklist for the user to follow.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Feb 17, 2026, 06:28 PM