system
Pass
Audited by Gen Agent Trust Hub on Nov 15, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill provides instructions for the agent to use headless browser tools (such as Chrome or Playwright) to render local HTML files and take screenshots for visual design review.
- Evidence: In
references/adapters/html.md, a command template is provided:"<chrome>" --headless=new --screenshot=board-1280.png --window-size=1280,900 file://$PWD/board.html. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted project data, including code, design briefs, and screen definitions, which presents a surface for potential prompt injection via embedded instructions.
- Ingestion points: Discovery and audit phases in
SKILL.mdandreferences/audit-checklist.mdrequire the agent to read existing project files, code, and design assets. - Boundary markers: The skill enforces multiple "Stop and show the person" checkpoints and explicit warnings in
SKILL.mdandreferences/audit-checklist.mdthat audits must not result in automatic fixes. - Capability inventory: The skill relies on file system access (read/write) and shell execution for rendering purposes.
- Sanitization: There is no programmatic sanitization of external file content, but the workflow is strictly gated by human review of all proposals and findings.
Audit Metadata