The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted search results from ScienceDirect without validation.\n
Ingestion points: The evaluate_script tool call in SKILL.md (Step 3) extracts paper titles, authors, and other metadata directly from the live page DOM.\n
Boundary markers: Absent; there are no instructions to the agent to treat the result content as untrusted or to ignore embedded directives within the scraped metadata.\n
Capability inventory: The skill utilizes navigate_page, evaluate_script, take_snapshot, and click via the Chrome DevTools MCP.\n
Sanitization: Absent; the extracted text is returned directly to the agent's context without any filtering or escaping.\n- [PROMPT_INJECTION]: The skill explicitly directs the agent to bypass security and bot-detection filters on the target website.\n
Evidence: The instructions mandate the use of an initScript to hide WebDriver status and provide a multi-step procedure to automatically detect and click the Cloudflare Turnstile captcha checkbox using a11y tree snapshots.\n- [DYNAMIC_EXECUTION]: The skill uses evaluate_script to execute custom JavaScript code for DOM interaction and data extraction.\n
Evidence: Step 3 defines a complex asynchronous JavaScript function used to poll for results and scrape metadata from the page context.

sd-search