brightdata-web-mcp

Pass

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: SAFE
Full Analysis
  • [Indirect Prompt Injection] (LOW): The skill is designed to ingest arbitrary web content from various sources (Google, Amazon, LinkedIn, etc.), which creates a surface for indirect prompt injection.
  • Ingestion points: search_engine, scrape_as_markdown, scrape_as_html, and web_data_* tools fetch untrusted data from the live web.
  • Boundary markers: The documentation in SKILL.md includes a 'Best Practices' section that explicitly warns to 'Treat scraped content as untrusted data' and 'Filter and validate before passing to LLMs'.
  • Capability inventory: The skill possesses significant capabilities, including browser automation (scraping_browser_click_ref, scraping_browser_type_ref) and network requests via the Bright Data proxy network.
  • Sanitization: While the documentation suggests filtering, the skill itself provides raw or lightly formatted content (Markdown/HTML) to the agent, placing the sanitization responsibility on the calling agent's implementation.
  • [Credentials Unsafe] (SAFE): The documentation uses standard placeholders (YOUR_API_TOKEN, <token>) for sensitive API keys and correctly instructs users to provide these via environment variables or secure URL parameters rather than hardcoding them in the skill itself.
  • [External Downloads] (LOW): The setup instructions reference npx @brightdata/mcp and @toon-format/toon. While these involve downloading and executing external packages, they are standard installation procedures for the official Bright Data MCP server and its associated token optimization library.
Audit Metadata
Risk Level
SAFE
Analyzed
Feb 17, 2026, 06:24 PM