The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/office/soffice.py performs runtime compilation of a C source file and uses the LD_PRELOAD environment variable to inject the resulting shared object library into the soffice (LibreOffice) process. This is designed to facilitate functionality in environments with Unix socket restrictions but utilizes advanced process injection techniques.
Evidence: Found subprocess.run(["gcc", "-shared", ...]) to compile a shim and env["LD_PRELOAD"] = str(shim) to inject it within scripts/office/soffice.py.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PowerPoint (.pptx) files and uses extracted data to drive agent actions, representing an indirect injection surface.
Ingestion points: Untrusted files enter the system via scripts/office/unpack.py and the markitdown tool mentioned in SKILL.md.
Boundary markers: The skill relies on standard XML tags and the OpenXML structure as delimiters for extracted content.
Capability inventory: The skill can execute multiple system commands via subprocess.run, including soffice, pdftoppm, gcc, and git across various helper scripts.
Sanitization: The skill uses defusedxml and lxml for XML parsing, providing protection against common XML-based attacks like entity expansion.

pptx