fig1-draw
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external research manuscripts and software repositories to extract scientific relationships, creating an inherent attack surface for indirect prompt injection.
- Ingestion points: Processes manuscipt content (PDF, LaTeX, Markdown) and source code repositories (SKILL.md).
- Boundary markers: Uses FigurePlan and FigureSpec JSON structures to isolate the extraction of scientific topology from the rendering process, preventing data from flowing directly into prompts without validation.
- Capability inventory: Invokes local Python scripts for validation and file manipulation, and uses image generation tools.
- Sanitization: Implements a specific utility script (clean_image_metadata.py) to strip all EXIF, XMP, and AI provenance markers from generated images prior to delivery.
- [DYNAMIC_EXECUTION]: The script scripts/build_overlay.py utilizes the import function to verify the runtime environment.
- Evidence: The script dynamically checks for the presence of hardcoded library names (PIL, pptx, vtracer, fitz) to ensure dependencies are available before proceeding with PPTX construction.
- [COMMAND_EXECUTION]: The skill executes local scripts and system utilities to perform validation and artifact generation.
- Evidence: The agent is instructed to run validation scripts (validate_figure_spec.py, validate_render_audit.py) to verify the integrity of generated specifications.
- Evidence: The build_overlay.py script executes the fc-list system command to verify available system fonts for text rendering in PowerPoint files.
Audit Metadata