image-to-code
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes several local commands and scripts to validate the environment and process design data. These include 'ls' for environment checks, 'python3' for design system searches and contrast validation, and 'node'/'npm' for rendering measurement and final verification. All scripts are located within the relative project directory structure.
- [INDIRECT_PROMPT_INJECTION]: The skill processes external visual references, such as screenshots and mockups, to infer design system properties. This data ingestion creates a potential surface for indirect prompt injection if the visual content contains hidden instructions meant for the vision model.
- Ingestion points: Visual references (images, screenshots, or mockups) provided by the user.
- Boundary markers: The skill provides a structured framework for the agent to analyze the image (Palette, Type, Spacing, Radius, Layout), which helps focus the model on specific attributes, though no formal delimiters for data extraction are used.
- Capability inventory: The skill allows for local file reading/writing, execution of Python and Node.js scripts, and running npm commands within the project context.
- Sanitization: The instructions explicitly direct the agent to avoid reproducing copyrighted imagery or brand assets, substituting them with generic placeholders, which serves as a task-level guardrail.
Audit Metadata