The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted web content which could contain malicious instructions intended to manipulate the agent's behavior.\n
Ingestion points: The skill fetches content from arbitrary URLs via the Chrome DevTools MCP, which is then processed by the Readability.js library and Markdown conversion scripts.\n
Boundary markers: The instructions in SKILL.md explicitly direct the agent to treat extracted text only as data and to ignore any embedded commands, requests for keys, or tool call suggestions.\n
Capability inventory: The skill possesses the ability to navigate the browser, download images, and write files (Markdown and JSON) to the local filesystem.\n
Sanitization: The extraction logic automatically removes potentially executable tags like <script>, <style>, and <iframe>. Content is further sanitized during the conversion to Markdown.\n- [EXTERNAL_DOWNLOADS]: The skill performs network requests to navigate web pages and download remote images.\n
Evidence: scripts/save_with_images.js and scripts/remote_url_policy.js handle outbound requests for navigation and asset collection.\n
Mitigation: The skill implements significant SSRF protections, including IP address blacklisting for private, loopback, and reserved ranges, DNS rebinding protection by pinning resolved IPs, and recursive validation of redirect URLs.\n- [DYNAMIC_EXECUTION]: The skill executes local JavaScript files within the browser context to perform extraction.\n
Evidence: SKILL.md defines a workflow where Readability.js and readability_extractor.js are loaded and run using the evaluate_script tool provided by the Chrome DevTools MCP.

web-article-extractor