visual-verification
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of rendered screenshots, DOM structures, and accessibility trees from arbitrary websites and applications.
- Ingestion points: Rendered web pages, application windows, and generated image files (SKILL.md).
- Boundary markers: None identified for visual data; the model relies on its internal vision capabilities to interpret pixels.
- Capability inventory: The skill uses browser automation (Playwright), Chrome DevTools MCP, and local PowerShell scripts to capture and analyze data.
- Sanitization: No sanitization is described for visual or structural data ingestion.
- [COMMAND_EXECUTION]: The skill instructs the agent to execute a diagnostic PowerShell script located at
TOOLS\Test-VisionCanary.ps1to validate its vision capabilities.
Audit Metadata