academic-figure-designer
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted natural-language architecture descriptions from users to generate specifications and natural-language prompts for image diffusion models.
- Ingestion points: User-provided text describing architecture or figure requirements is ingested through a normalization protocol and stored within the sources metadata field of the FigureSpec.
- Boundary markers: The skill enforces strict prose normalization and explicitly prohibits markdown syntax in generated prompts, which serves to prevent image models from misinterpreting formatting as control instructions.
- Capability inventory: The skill uses local Python scripts for validation and sanitization tasks and interfaces with a platform-native image generation tool.
- Sanitization: The skill includes validate_figure_spec.py to enforce path safety by preventing symlink attacks and ensuring all operations occur within a trusted workspace root. It also includes clean_image_metadata.py to strip AI-generation metadata (such as EXIF and C2PA) from output images. These risks are inherent to the skill's primary function and are mitigated through robust design practices.
Audit Metadata