The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill installs the peekaboo CLI tool using the Homebrew package manager from a third-party repository (steipete/tap/peekaboo). This introduces an external dependency that is not managed by the skill vendor or the official Homebrew core repository.
[DATA_EXFILTRATION]: The skill can access sensitive user data through several commands: peekaboo clipboard reads the contents of the macOS clipboard, while peekaboo see, peekaboo image, and peekaboo capture allow the agent to view or record the user's screen.
[COMMAND_EXECUTION]: The skill executes high-privilege UI automation commands that can control the entire macOS interface, including simulating keyboard and mouse input (click, type, press), managing applications (app launch/quit), and executing automation scripts via peekaboo run. It also manages provider settings and credentials through the peekaboo config command.
[PROMPT_INJECTION]: The skill's architecture is vulnerable to indirect prompt injection because it interprets visual data from the screen. * Ingestion points: Untrusted text from the user's screen is ingested by the agent through UI snapshots (peekaboo see) and screen captures. * Boundary markers: There are no explicit markers or instructions provided to the agent to help it differentiate between legitimate instructions and text found within captured screenshots or UI maps. * Capability inventory: The skill possesses powerful capabilities such as script execution (run), input simulation (type, click), and the ability to launch applications, which could be abused if the agent follows instructions found on a malicious webpage or document. * Sanitization: No sanitization or filtering of the captured UI text is performed before it is passed to the AI model for interpretation.

peekaboo