The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/blind.mjs initiates a local HTTP server to host the comparison interface.
Evidence: http.createServer(...).listen(port, ...) in scripts/blind.mjs.
[DYNAMIC_EXECUTION]: The web interface uses innerHTML to render folder names and results, which allows for script execution (XSS) if folder names contain malicious HTML tags.
Evidence: The render function in the PAGE constant within scripts/blind.mjs uses main.innerHTML to display rows, which are derived from the directory names provided as command-line arguments.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted local files (HTML, images, videos) and displays them to the user. Malicious files or directory names could be used to execute scripts or influence the agent's behavior.
Ingestion points: Folder names (dirA, dirB) and design files within those folders.
Boundary markers: None for the folder names rendered in the final summary view.
Capability inventory: The script can read any file in the specified directories, write to the results JSON file, and start a local network server.
Sanitization: Absent for directory names rendered in the UI; brief content uses textContent but designs are served via iframes as-is.

kero-blind