The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill installs the peekaboo binary via a third-party Homebrew tap (steipete/tap/peekaboo). This introduces a dependency on an external source that is not listed as a trusted vendor.
[COMMAND_EXECUTION]: Provides the agent with full macOS UI automation capabilities, including launching/quitting applications (app), executing scripts (run), and performing hardware-level interactions like clicks, drags, and key presses.
[DATA_EXFILTRATION]: Grants access to potentially sensitive screen content and user data through live capture, screenshots, and annotated UI maps (capture, image, see). The clipboard command also allows the agent to read the system clipboard, which often contains passwords or private data.
[CREDENTIALS_UNSAFE]: The config command explicitly supports managing 'credentials', which could lead to the exposure or manipulation of API keys and authentication tokens stored within the tool's configuration.
[PROMPT_INJECTION]: The skill is highly vulnerable to indirect prompt injection. Because it captures and analyzes screen content, an attacker could display malicious instructions on a website or in a document that the agent 'sees' and mistakenly follows.
Ingestion points: Screen captures, UI element inspection (see), and clipboard reading.
Boundary markers: None identified; raw UI content is passed to the agent's context.
Capability inventory: Full OS-level interaction (click, type, open URLs, manage apps).
Sanitization: No evidence of filtering or sanitizing captured text or UI labels before processing.

peekaboo