The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Vulnerability to command injection in pywebview/window.py. The show_notification function constructs an AppleScript command via f-strings using the message and title parameters provided in dialog requests. Since these inputs are not sanitized, a malicious payload (e.g., containing double quotes and the do shell script command) can break out of the string context and execute arbitrary shell commands on the host system.
[COMMAND_EXECUTION]: The script tool.js makes extensive use of child_process.spawn and child_process.execSync to manage its environment. This includes starting the Express backend, invoking Python for window management, and executing build scripts like npm run build.
[EXTERNAL_DOWNLOADS]: During installation or updates, tool.js triggers npm install in multiple directories, which downloads third-party packages from the public NPM registry.
[COMMAND_EXECUTION]: The skill launches pywebview/window.py as a detached subprocess to manage native system windows, as seen in the startPywebview function in tool.js.

agent-interact