The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted visual data from the user's screen which may contain adversarial instructions designed to influence the agent's behavior.
Ingestion points: Screen capture data processed by the vu_describe tool as described in SKILL.md.
Boundary markers: None identified in the skill instructions to delimit screen content from system instructions.
Capability inventory: The skill possesses extensive GUI automation capabilities including vu_click, vu_type, vu_hotkey, vu_move, vu_drag, and vu_scroll (defined in SKILL.md).
Sanitization: No evidence of sanitization or filtering of text extracted from screen captures before it is processed by the AI vision models.
[DATA_EXFILTRATION]: The skill is designed to transmit screen capture data to external AI service providers (Claude, OpenAI, Gemini) for analysis. While this is the primary functionality, it involves sending potentially sensitive visual information from the user's desktop to remote servers.
[COMMAND_EXECUTION]: The skill provides tools for direct GUI interaction and automation, allowing the agent to simulate mouse clicks, keyboard typing, and system hotkeys. While safety blocks are mentioned for specific hotkeys (e.g., cmd+q), the ability to automate arbitrary UI actions represents a significant control surface.

omni-vu