The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/thumbnail.py executes external system commands (libreoffice and pdftoppm) via subprocess.run to render slide previews. The implementation uses a list-based argument structure which prevents shell injection.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided PowerPoint files (.pptx) which serves as an ingestion point for untrusted data. A malicious deck could contain instructions intended to influence the agent's behavior during inspection or exploit vulnerabilities in processing tools like LibreOffice.
Ingestion points: scripts/thumbnail.py (pptx file), scripts/add_slide.py (unpacked directory), and scripts/clean.py (unpacked directory).
Boundary markers: The SKILL.md file includes an OWASP ASI guardrail section advising the agent to treat imported content as data, though no programmatic delimiters or instruction filtering are enforced in the scripts themselves.
Capability inventory: scripts/thumbnail.py uses subprocess.run to execute system binaries. scripts/add_slide.py and scripts/clean.py perform file reads, writes, and deletions using standard library functions.
Sanitization: Path handling uses the pathlib module which mitigates basic path traversal. XML processing uses the standard xml.etree.ElementTree library.

pptx