The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides templates for executing shell commands via the bash tool and subprocess.run, as well as system-level interactions using pyautogui. This design provides a broad capability set that requires the strict sandboxing practices described in the documentation to ensure safety.
[INDIRECT_PROMPT_INJECTION]: By design, the skill captures and processes screen content using vision-language models. This creates an attack surface where untrusted visual data (such as web pages or third-party applications) could contain instructions meant to override the agent's logic.
Ingestion points: Screen captures processed in SKILL.md.
Boundary markers: Not explicitly implemented in the provided code snippets.
Capability inventory: Full mouse and keyboard control via pyautogui and shell access via the bash tool.
Sanitization: Not present in the architectural patterns provided.

computer-use-agents