skills/trycua/cua/jev-use/Gen Agent Trust Hub

jev-use

Pass

Audited by Gen Agent Trust Hub on Sep 20, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [SAFE]: The skill documents a design pattern for integrating an agent with UI automation tools. It focuses on defining a secure communication protocol (stdin/stdout) and decision boundaries between components.
  • [COMMAND_EXECUTION]: The instructions describe invoking a Python interpreter and external scripts. To mitigate security risks, it explicitly mandates using absolute paths and direct invocation without a shell, which prevents common shell injection vulnerabilities.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and act upon untrusted data from the environment, such as browser DOM and accessibility trees. It incorporates multiple defense-in-depth measures:
  • Ingestion points: Browser DOM, accessibility tokens, and visual regions provided by Cua Driver (SKILL.md).
  • Boundary markers: Use of a 'bounded candidate table' where the application, not the model, defines valid actions; the model only selects an ID.
  • Capability inventory: Execution of a single Driver action after validation; background delivery by default.
  • Sanitization: Mandatory validation of capture IDs, coordinate mapping, and dimensions. The skill requires rejecting malformed, stale, or unknown choices and verifying postconditions independently of the model's response.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 20, 2026, 03:27 AM
Security Audit — agent-trust-hub — jev-use