The Agent Skills Directory

[DATA_EXFILTRATION]: The screenshot and recording engine in scripts/shoot.mjs implements a mandatory path filter (privatePath) that blocks the browser from accessing sensitive directories and files, including .ssh, .aws, .env, and files with names like credentials or secrets. This prevents rendered pages from harvesting local secrets.
[REMOTE_CODE_EXECUTION]: The version update utility in scripts/check_update.py executes a command-line tool hosted on the author's official GitHub repository. The execution is cryptographically pinned to a specific commit hash (64ada13fbf95d3154fa708cedaa1b3a963dc1151), ensuring that only the audited version of the updater can run.
[DYNAMIC_EXECUTION]: Portable design comparison pages generated by scripts/build_explorer.py are protected by a strict Content Security Policy. Previews are sandboxed with connect-src 'none' and script-src 'none' (or unsafe-inline for interactive prototypes), effectively preventing design samples from connecting to external servers.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted HTML and CSS data during the design review phase. It mitigates injection risks by performing all rendering in an isolated browser session and sanitizing metadata in the build_explorer.py script to prevent script tag escapes.
[EXTERNAL_DOWNLOADS]: The scripts/curate-media.mjs script fetches assets from established cultural and educational organizations (e.g., NASA, the Library of Congress, and the Met Museum). These operations are performed via standard HTTP requests, and the tool maintains a SOURCES.md log to ensure transparency of asset origins and licensing.

oil-ui