The Agent Skills Directory

[DATA_EXFILTRATION]: The HTML templates taobao_main_3x4.html and taobao_main_800x800.html contain hardcoded image sources using the file:// protocol to access /root/.openclaw/workspace/model.jpg. Accessing the /root/ directory and internal application folders (like .openclaw) is a significant security risk, indicating an attempt to harvest data from sensitive system locations or configurations.\n- [COMMAND_EXECUTION]: The instructions in SKILL.md generate shell commands using unvalidated user input for parameters such as product names, selling points, and prices. This pattern creates a potential command injection surface if the AI agent executes crafted input within the shell environment.\n- [EXTERNAL_DOWNLOADS]: The skill documentation encourages the execution of .venv/bin/playwright install chromium, which downloads and installs binary executables from remote servers. While Playwright is a well-known service, the automated fetching of binaries increases the risk profile when combined with other suspicious file access patterns.\n- [PROMPT_INJECTION]: The render_template.py script performs a direct string replacement of user input into HTML templates that are then rendered by a headless browser. This lack of sanitization allows for Indirect Prompt Injection (HTML/JS injection), which could be used to facilitate Server-Side Request Forgery (SSRF) or further local file disclosure by injecting malicious tags into the rendering engine.

ecommerce-designer