design-screenshot

Pass

Audited by Gen Agent Trust Hub on May 11, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection by processing external data (screenshots) provided via the $ARGUMENTS placeholder. This data is analyzed by multimodal vision tools (ai-multimodal) to guide the creation of design plans and implementation of code. An attacker could provide a screenshot containing embedded text instructions designed to bypass agent constraints or inject malicious logic into the resulting frontend code.
  • Ingestion points: SKILL.md (via <screenshot>$ARGUMENTS</screenshot>)
  • Boundary markers: Absent. There are no explicit instructions to the vision model to ignore text found within the image that might conflict with system prompts.
  • Capability inventory: The skill can create directories, write files (plan.md, .md phase files, ./docs/design-guidelines.md), and use subagents to implement HTML/CSS/JS code.
  • Sanitization: Not detected. The skill relies on the AI's vision capabilities to interpret the input without a validation or sanitization layer for instructions found within the image.
  • [PROMPT_INJECTION]: The skill uses repetitive, high-emphasis directives such as 'MANDATORY IMPORTANT MUST ATTENTION' and role-play instructions ('ALWAYS REMEMBER that you have the skills of a top-tier UI/UX Designer'). These are characteristic of prompt engineering techniques used to steer agent behavior and override default constraints, though here they appear intended for quality assurance rather than malicious bypass.
Audit Metadata
Risk Level
SAFE
Analyzed
May 11, 2026, 05:52 AM
Security Audit — agent-trust-hub — design-screenshot