markitdown
Pass
Audited by Gen Agent Trust Hub on Mar 31, 2026
Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The skill contains steering instructions that direct the agent to always generate scientific diagrams and use a specific external skill ('scientific-schematics') by default, which attempts to override the agent's natural decision-making process.
- [COMMAND_EXECUTION]: Instructions direct the agent to execute a local script 'scripts/generate_schematic.py' which is not included in the skill's distribution, potentially leading the agent to attempt unverified actions or synthesize code.
- [EXTERNAL_DOWNLOADS]: Fetches the 'markitdown' library from Microsoft's official GitHub and PyPI repositories and utilizes the OpenRouter API for AI-enhanced image descriptions.
- [PROMPT_INJECTION]: The skill processes untrusted external data (PDF, DOCX, XLSX, etc.) and converts it to Markdown for LLM consumption, creating an indirect prompt injection surface. Ingestion points: Document conversion methods in SKILL.md and provided Python scripts; Boundary markers: No delimiters or warnings are used to isolate converted content; Capability inventory: The skill is granted Bash, Read, Write, and Edit tools; Sanitization: No escaping or validation is performed on the text extracted from documents.
Audit Metadata