The Agent Skills Directory

DATA_EXFILTRATION (MEDIUM): The vu_describe tool captures the user's screen and transmits image data to third-party AI providers (Claude, OpenAI, Gemini) for processing. This presents a risk of exposing sensitive information visible on the display, such as credentials, private communications, or internal documents.
PROMPT_INJECTION (LOW): The skill is susceptible to indirect prompt injection via visual content. Ingestion points: Screen captures processed by vu_describe. Boundary markers: None; the AI interprets the raw visual scene. Capability inventory: Full GUI control including vu_click, vu_type, and vu_hotkey. Sanitization: None; the vision model attempts to interpret all visible text as potential context or instruction.
COMMAND_EXECUTION (MEDIUM): Automation tools like vu_type and vu_hotkey allow the agent to simulate keyboard input and system shortcuts. This can be used to execute arbitrary shell commands (e.g., by opening a terminal) if the agent is manipulated via prompt injection.
PRIVILEGE_ESCALATION (MEDIUM): The skill requires the user to grant high-privilege macOS permissions: 'Screen Recording' and 'Accessibility'. These permissions provide the agent with broad visibility and control over the operating system environment.

omni-vu