webclaw
Pass
Audited by Gen Agent Trust Hub on Jul 23, 2026
Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The skill possesses an indirect prompt injection attack surface as it is designed to fetch and process untrusted data from the web for AI agent consumption.
- Ingestion points: The scrape, crawl, map, batch, extract, and lead tools all fetch content from external URLs provided at runtime.
- Boundary markers: The skill implements an llm output format that includes structural headers to help agents distinguish content from metadata, though this is not a complete defense against adversarial content.
- Capability inventory: The skill can perform network requests, search the web via APIs, and process content through local or cloud-based LLMs.
- Sanitization: The tool converts raw HTML into cleaned Markdown, text, or JSON, which reduces the complexity of the input but does not filter semantic instructions.
- [EXTERNAL_DOWNLOADS]: The skill utilizes the Node package manager to download and execute its execution engine.
- The npx create-webclaw command is used for initial setup and configuration.
- The npx @webclaw/mcp command is used to fetch and run the Model Context Protocol server.
Audit Metadata