The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run in scripts/tiers/tier2_5_agentbrowser.py and scripts/session/manager.py to execute the agent-browser CLI. This involves passing arguments derived from input URLs and session states to the system shell.
[EXTERNAL_DOWNLOADS]: The implementation of Tier 2 (cloakbrowser) and Tier 3 (camoufox) triggers the download of large, unverified browser binary packages (~200MB and ~780MB respectively) from external registries during execution.
[REMOTE_CODE_EXECUTION]: All browser-based tiers (scripts/tiers/tier2_scrapling.py, scripts/tiers/tier2_5_agentbrowser.py, and scripts/tiers/tier3_camoufox.py) use dynamic code execution via page.evaluate() or eval() to inject the skill's own logic (e.g., Shadow DOM piercing, WebMCP interception) into the browser context of the target website.
[PROMPT_INJECTION]: The skill is vulnerable to Indirect Prompt Injection (Category 8). It ingests arbitrary content from untrusted websites and interpolates it into prompts for the AI extraction router (scripts/extraction/ai_router.py) and vision router (scripts/extraction/vision_router.py) without sufficient boundary markers or 'ignore' instructions. This ingested data is processed by the agent which has capabilities including network access and file writes.

ultimate-scraper