image-to-editable-ppt
Pass
Audited by Gen Agent Trust Hub on Sep 13, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill integrates with external services for OCR (PaddleOCR via Baidu's AI Studio) and image generation (OpenAI/ChatGPT). These calls are essential for the primary purpose of converting images to editable PowerPoint slides. The skill instructions include explicit privacy safeguards, directing the agent not to send unrelated sensitive files or credentials to these third-party APIs.
- [COMMAND_EXECUTION]: The editppt CLI executes established local system tools via subprocesses, including LibreOffice (soffice) for document normalization, LaTeX engines (xelatex, pdflatex) for formula rendering, and ImageMagick (magick) for image processing. The LaTeX renderer includes an optional shell-escape flag, which is documented as a risk and disabled by default. These are legitimate uses within the scope of a document reconstruction tool.
- [DYNAMIC_EXECUTION]: The CLI architecture uses runpy.run_path to load its main runtime and subprocess.run to coordinate various internal Python scripts (prepare_deck_run.py, build_pptx_from_manifest.py, etc.). This is a standard method for organizing multi-step CLI workflows.
- [INDIRECT_PROMPT_INJECTION]: The skill has a surface for indirect prompt injection as it processes untrusted user documents. 1. Ingestion points: editppt prepare reads user-provided images, PDFs, and PPTX files. 2. Boundary markers: The skill uses explicit worker prompts generated from prompts/page-worker.md which define task boundaries and ownership. 3. Capability inventory: The skill can execute local commands for rendering and call external image generation tools. 4. Sanitization: The skill provides structural validation of manifests and instructs agents to manually verify OCR text against visual sources to prevent obedience to embedded instructions.
Audit Metadata