gemini-search-mcp-web-search

Pass

Audited by Gen Agent Trust Hub on Sep 16, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill implements tools (web_search and ask) that fetch and synthesize content from the live web, introducing a surface for indirect prompt injection attacks.
  • Ingestion points: The web_search(query) and ask(prompt) tools ingest untrusted data from external websites.
  • Boundary markers: The documentation does not specify the use of delimiters or 'ignore instructions' directives for the fetched content.
  • Capability inventory: The skill possesses network access, browser control via subprocesses, and the ability to process and synthesize web data for the agent.
  • Sanitization: There is no mention of sanitizing or filtering instructions that might be embedded in the fetched search results.
  • [DATA_EXFILTRATION]: The skill encourages the use of persistent browser profiles via the GEMINI_SEARCH_USER_DATA_DIR environment variable to bypass CAPTCHAs.
  • Evidence: The configuration instructions suggest pointing the server to persistent directories like ~/.local/share/gemini-search-mcp/chrome-profile. Such profiles typically contain sensitive information including cookies, session tokens, and browsing history which could be exposed if the server or search session is compromised.
  • [COMMAND_EXECUTION]: The server manages and launches browser processes using local command execution.
  • Evidence: The GEMINI_SEARCH_BROWSER_BACKEND defaults to subprocess, and the documentation provides instructions for launching various browser channels (Chrome, Edge, Chromium) directly from the host system.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 16, 2026, 10:44 PM
Security Audit — agent-trust-hub — gemini-search-mcp-web-search