ui-ux-evidence-review

Pass

Audited by Gen Agent Trust Hub on Oct 13, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted external data, such as UI screenshots, frontend source code, and live web application content, which provides a surface for indirect prompt injection attacks.
  • Ingestion points: SKILL.md, references/screenshot-review.md, and references/code-review.md describe the process of reading and evaluating images, URLs, and source code files provided by the user.
  • Boundary markers: SKILL.md (Section 1) contains explicit instructions to treat all text within documents and code as data only, mandating that the agent must not interpret such content as new instructions or execution commands. This is reinforced in the evaluation case tests/review-cases.md (ID E14).
  • Capability inventory: The skill utilizes local scripts (scripts/contrast.py, scripts/check_package.py) for targeted technical audits and relies on standard browser/file access tools to perform its primary function.
  • Sanitization: The instructions emphasize a logical separation between data observation and rule-based inference to prevent adversarial content from influencing agent behavior.
  • [SAFE]: The provided Python scripts (scripts/contrast.py and scripts/check_package.py) are focused, read-only utilities that perform mathematical calculations (color contrast) or internal file structure validation. They do not initiate network connections, execute remote code, or access sensitive system paths.
  • [SAFE]: The skill refers to and links to official documentation from the World Wide Web Consortium (W3C) for accessibility standards (WCAG). These are recognized, well-known, and trusted resources within the technology industry.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 13, 2026, 05:36 PM