The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the pyautogui library to implement GUI control, granting an agent the ability to simulate mouse clicks, keyboard typing, and scrolling on the host system.\n- [REMOTE_CODE_EXECUTION]: The pattern includes a 'bash' tool implementation that uses the subprocess module to execute arbitrary shell commands. This is a highly privileged capability that requires strict sandboxing as recommended in the documentation.\n- [DATA_EXFILTRATION]: The skill uses pyautogui.screenshot() and the scrot utility to capture screen images. These images, which may contain sensitive personal or system data, are encoded and transmitted to the AI model for processing.\n- [PROMPT_INJECTION]: As a vision-enabled agent that interacts with external interfaces (like web browsers), it is susceptible to indirect prompt injection where malicious content on a screen could manipulate the agent's behavior.\n
Ingestion points: Screen captures processed in the perception loop (e.g., capture_screenshot).\n
Boundary markers: None implemented in the provided template code; the agent processes raw desktop images.\n
Capability inventory: Desktop control via pyautogui and shell command execution via the bash tool.\n
Sanitization: No input validation or filtering is applied to the screen data before processing.

computer-use-agents