The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the pyautogui library to control the host system's mouse and keyboard (clicking, typing, scrolling) and the subprocess module to execute shell commands like scrot for screen capture.
[EXTERNAL_DOWNLOADS]: The included Dockerfile specifies the installation of various system utilities (xvfb, x11vnc, fluxbox, xterm, firefox, etc.) from official Ubuntu repositories and Python packages from standard registries.
[PROMPT_INJECTION]: The skill architecture creates a surface for indirect prompt injection by processing untrusted visual data from screenshots. 1. Ingestion points: The capture_screenshot method in SKILL.md captures the screen state for model analysis. 2. Boundary markers: No explicit text-based boundary markers are used for visual input. 3. Capability inventory: The skill implements pyautogui for desktop control and a bash tool for shell command execution. 4. Sanitization: No input sanitization is present, as the skill relies on architectural sandboxing for security.
[SAFE]: The skill serves as a security-conscious guide for high-privilege agent behaviors. It includes explicit instructions and ready-to-use configurations (Dockerfile, Docker Compose) for sandboxing, privilege reduction (non-root users), and network isolation to contain the risks associated with computer-use agents.

computer-use-agents