The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implements desktop automation using the pyautogui library, allowing the agent to perform actions such as clicking, typing text, and pressing keys on the host or virtual system.
[DYNAMIC_EXECUTION]: The implementation includes tool definitions for a vision-language model to execute shell commands (bash) and perform file system operations via a text editor, which involves runtime command generation and execution.
[INDIRECT_PROMPT_INJECTION]: The Perception-Reasoning-Action loop ingests screenshots of the desktop environment to drive agent decisions. This architecture is susceptible to indirect prompt injection if the agent processes screens containing malicious instructions (e.g., from a browser window or untrusted document).
Ingestion points: pyautogui.screenshot() in SKILL.md provides visual input of the current desktop state to the model.
Boundary markers: No specific prompt delimiters or instructions to ignore embedded visual commands are implemented in the provided logic.
Capability inventory: The skill exposes full GUI control (pyautogui) and shell command execution (subprocess) capabilities to the agent.
Sanitization: There is no evidence of filtering, validation, or sanitization of the visual data before it is interpreted by the reasoning model.

computer-use-agents