The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides a standard Docker command for execution in SKILL.md. The command uses the '--rm' flag for environment cleanup and a specific volume mount restricted to the local assets directory, preventing the container from accessing sensitive host system paths.
[PROMPT_INJECTION]: The skill includes extensive defensive guardrails in both SKILL.md and SAFETY.md. These instructions explicitly harden the agent against prompt injection and social engineering attempts, instructing the agent to ignore user claims of authority or urgency when they conflict with security boundaries like local-only data processing and Video ID verification.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it ingests data from external websites (YouTube, X, and generic URLs) and processes it for agent output. Evidence: 1. Ingestion points: Web content is retrieved via the PlaywrightCrawler in assets/main_handler.js and assets/youtube_handler.js. 2. Boundary markers: SKILL.md includes explicit constraints to ignore embedded instructions in scraped data and maintain local processing boundaries regardless of external claims. 3. Capability inventory: The skill performs containerized command execution and prints data to stdout for agent consumption. 4. Sanitization: The JavaScript handlers strip HTML/XML tags and truncate the resulting text to 15,000 characters to reduce the risk of large-scale payload delivery.
[EXTERNAL_DOWNLOADS]: The skill utilizes standard, well-known Node.js packages 'crawlee' and 'playwright' for its core functionality. These dependencies are listed in package.json and do not involve the execution of untrusted remote code or scripts at runtime.

deep-scraper-hardened