The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/office/soffice.py contains a sophisticated shimming mechanism. It generates a C source file at runtime, compiles it into a shared object using gcc, and then injects it into the LibreOffice process via the LD_PRELOAD environment variable. This is used to intercept and redirect network-related system calls to circumvent sandbox restrictions.
[COMMAND_EXECUTION]: Multiple scripts invoke system binaries using the subprocess module. scripts/office/soffice.py executes gcc for building the library shim; scripts/thumbnail.py executes pdftoppm to generate slide thumbnails; and scripts/office/validators/redlining.py executes git diff to compare text content across document versions.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted PowerPoint files, which creates an attack surface for indirect prompt injection.
Ingestion points: Natural language text is extracted from slides using the markitdown tool as described in SKILL.md.
Boundary markers: The skill uses  markdown comments to separate content blocks.
Capability inventory: Scripts in the skill can execute arbitrary system commands via subprocess.run (e.g., soffice, pdftoppm).
Sanitization: While the skill uses defusedxml to prevent XML-based attacks, it does not apply specific sanitization or filtering to the natural language content extracted from presentation slides before it enters the agent's context.
[SAFE]: The skill follows several security best practices, such as using the defusedxml library in scripts/clean.py and scripts/office/validators/base.py to mitigate XML External Entity (XXE) risks. Additionally, the safe_extract utility in scripts/office/helpers/__init__.py provides protection against path traversal (ZipSlip) vulnerabilities during file extraction.

pptx