The Agent Skills Directory

[COMMAND_EXECUTION] (LOW): Based on the unit tests in scripts/tests/test_document_converter.py, the skill uses subprocess.run to execute external binaries such as wkhtmltopdf for HTML and Markdown conversion.
Evidence: Functions like test_convert_markdown_success and test_convert_html_success mock subprocess calls, indicating that the underlying document_converter.py performs system-level command execution.
Risk: If input filenames or content are not properly sanitized before being passed to the shell, this could lead to command injection.
[PROMPT_INJECTION] (LOW): The skill is designed to ingest and process untrusted external data (Audio, PDF, Word, and HTML files).
Ingestion Points: audio-processing.md (audio files via client.files.upload), test_document_converter.py (PDF, Markdown, HTML inputs).
Boundary Markers: Absent. The prompts provided in references/audio-processing.md (e.g., "Generate a transcript of the speech.") do not use delimiters or instructions to ignore embedded commands.
Capability Inventory: Subprocess execution, file system writes (open('output.wav', 'wb')), and network access to the Gemini API.
Sanitization: No sanitization or validation of the input file content is visible in the provided reference code or tests.

ai-multimodal