academic-figure-workflow
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to execute local Python scripts (
validate_figure_spec.pyandclean_image_metadata.py) to validate figure specifications and sanitize image artifacts. These operations are performed usingpython3with structured arguments, and the instructions explicitly warn against interpolating user-controlled labels into shell commands to prevent injection attacks. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, including academic papers (PDFs, URLs), source code repositories, and reference images. Ingestion points: User-provided URLs for papers/repos, PDF files, and image files are read via the agent's browser or document reader tools. Boundary markers: The skill uses a structured FigurePlan and FigureSpec (JSON) to separate analysis from execution and includes a 'Missing-Info Policy' to label inferences clearly. Capability inventory: The skill can write image files to the workspace, execute specific local Python scripts for validation and cleaning, and call native image generation tools. Sanitization: It includes a mandatory step to run
clean_image_metadata.pyto strip potentially sensitive metadata (EXIF, C2PA) from generated images before delivery.
Audit Metadata