The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill's metadata specifies the installation of an external binary via a third-party Homebrew tap (steipete/tap/peekaboo). This binary is executed to perform all skill functions, representing the execution of unverifiable external code.
[COMMAND_EXECUTION]: The skill provides a comprehensive suite of commands to control the host system, including:
Simulating keyboard input (type, hotkey, press)
Mouse and gesture automation (click, drag, move, scroll, swipe)
Application and window management (app, window, dock, menu)
Executing arbitrary scripts via peekaboo run with JSON payloads.
[DATA_EXFILTRATION]: The skill facilitates the capture and potential exposure of sensitive information through multiple vectors:
peekaboo image and peekaboo see capture screen content, windows, and menu bars.
peekaboo clipboard reads and writes to the system clipboard.
peekaboo config manages credentials and model provider settings.
The see and image commands support an --analyze flag, which likely sends captured visual data to external LLM providers for processing.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection.
Ingestion points: Data is ingested from the macOS UI via peekaboo see, peekaboo image, and peekaboo list (SKILL.md).
Boundary markers: None present; there are no instructions to the agent to ignore or delimit instructions found within the captured UI data.
Capability inventory: Extensive system control capabilities including keystroke injection, file system access via open, and application launching (SKILL.md).
Sanitization: No evidence of sanitization or filtering of the text or visual data captured from the screen before it is processed by the agent.

peekaboo