The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides an indirect prompt injection surface by design, as its core function is to fetch and process content from arbitrary external URLs.
Ingestion points: External web content is ingested from URLs provided to the extraction tool (SKILL.md).
Boundary markers: The tool returns data in a structured JSON format containing a raw_content field, which provides a structural boundary between external data and the agent's instruction stream.
Capability inventory: The skill relies on shell execution (./scripts/extract.sh) and network operations via curl to api.tavily.com.
Sanitization: The instructions do not describe any sanitization, filtering, or safety checks performed on the retrieved content before it is returned to the agent context.
[DATA_EXFILTRATION]: The skill interacts with sensitive local file paths to manage authentication credentials.
Evidence: The skill is documented to check ~/.mcp-auth/ for OAuth tokens and can read API keys from ~/.claude/settings.json. While this is standard behavior for Model Context Protocol (MCP) authentication, it involves the programmatic access of credential storage locations.
[EXTERNAL_DOWNLOADS]: The skill performs outbound network requests to api.tavily.com to retrieve content. Tavily is a well-known service specifically designed for providing clean web data to AI agents.

extract