The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts scripts/convert.py and scripts/validate_output.py utilize subprocess.run to call external conversion tools such as markitdown, pandoc, and pdftotext. These implementations use argument lists, which is a secure practice to prevent shell injection. Evidence: subprocess.run(['markitdown', str(file_path)], capture_output=True, ...) in scripts/convert.py.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted document files (PDF, DOCX, etc.) which could contain adversarial instructions. If the agent processes the resulting Markdown without further safety checks, it could be influenced by embedded prompts. Evidence Chain: 1. Ingestion points: scripts/convert.py and scripts/validate_output.py. 2. Boundary markers: Absent. 3. Capability inventory: subprocess.run (tool invocation) and file system writes. 4. Sanitization: Absent for document content.

markdown-tools