computer-use

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill operates by capturing and interpreting screenshots and accessibility trees from the user's desktop, creating a surface where malicious content within a window or webpage could attempt to hijack the agent's logic.
  • Ingestion points: Screen captures and AX-tree accessibility indices retrieved via the computer_use(action="capture") tool (defined in SKILL.md).
  • Boundary markers: The skill contains explicit instructions for the agent to ignore directions found in screen content and to treat the user's prompt as the sole source of truth.
  • Capability inventory: The skill provides access to GUI automation (clicking and typing), terminal command execution, and file system operations (write_file, patch).
  • Sanitization: The skill documents tool-level filtering meant to block dangerous shell patterns (e.g., sudo rm -rf, curl | bash) during keyboard input simulation.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 03:48 PM
Security Audit — agent-trust-hub — computer-use