The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on executing system commands such as xdotool, scrot, imagemagick, and xdpyinfo via the Python subprocess module.
Evidence: The script scripts/cu.py uses subprocess.run to call these binaries for automation tasks. The implementation avoids shell=True and uses the -- separator for xdotool commands, which effectively mitigates standard command injection risks from user-supplied text or keys.
[PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection (Category 8) because it captures and returns screenshots of the desktop environment to the agent.
Ingestion points: Visual data enters the agent context via cu-screenshot and the take_screenshot function in scripts/cu.py.
Boundary markers: There are no boundary markers or 'ignore' instructions applied to the visual content rendered by the GUI applications.
Capability inventory: The skill possesses the capability to simulate keyboard and mouse input via xdotool, providing a mechanism for an injection to trigger further actions.
Sanitization: No sanitization is performed on the visual content before it is processed by the agent's vision capabilities. This is an inherent risk of visual automation where the agent might follow instructions displayed within a GUI app.

computer-use-cli