academic-figure-designer
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes natural language architecture descriptions and external reference images from users, which serves as a potential ingestion point for indirect instructions. This risk is mitigated by explicit boundary instructions (e.g., 'Zero Markdown Syntax') and a mandatory prompt review workflow defined in the spec.
- Ingestion points: User-provided natural language architecture descriptions and external reference images referenced in FigureSpec JSON.
- Boundary markers: The skill employs 'Negative Constraints' and prose normalization rules to prevent the model from interpreting structural markers as commands.
- Capability inventory: The agent is authorized to execute validation scripts, process images using Pillow, and call the image_gen.imagegen tool.
- Sanitization: All file operations are sanitized via validate_figure_spec.py, which validates that workspace_root and output_path are contained within a trusted directory to prevent path traversal.
- [COMMAND_EXECUTION]: The skill requires the execution of local Python scripts (validate_figure_spec.py and clean_image_metadata.py) to manage its workflow. These scripts are provided as part of the skill and perform defensive checks, such as verifying the integrity of the workspace and stripping sensitive metadata from generated artifacts.
Audit Metadata