macos-computer-use

Pass

Audited by Gen Agent Trust Hub on Jun 13, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill requires an external driver called cua-driver to be installed via its upstream script through the platform's tool manager (hermes tools). This is a necessary dependency for the stated GUI automation purpose.
  • [COMMAND_EXECUTION]: The skill enables comprehensive control over macOS GUI elements, including keyboard input (type, key) and application management (focus_app). The author includes documentation of a platform-level guard that blocks dangerous command patterns, such as sudo rm -rf or piped shell installations.
  • [DATA_EXFILTRATION]: The capture action allows the agent to take screenshots and read accessibility trees (AX-tree) of running applications. To mitigate the risk of credential harvesting, the skill includes explicit 'Hard Rules' forbidding the typing of passwords, API keys, or secrets, and warns the agent to avoid personal or sensitive application windows.
  • [PROMPT_INJECTION]:
  • Ingestion points: The skill ingests untrusted data from the macOS UI and web pages via the capture action (specifically the som and vision modes).
  • Boundary markers: The instructions include a specific 'Safety' section that commands the agent to treat the user's prompt as the only source of truth and to ignore any instructions found within captured screenshots or web content.
  • Capability inventory: The skill can perform mouse/keyboard interaction, drag-and-drop, and application focusing across the OS.
  • Sanitization: The skill documentation references a 'guard' at the tool level that prevents the execution of blocked patterns and system-level shortcuts.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 13, 2026, 01:28 AM
Security Audit — agent-trust-hub — macos-computer-use