The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/office/soffice.py generates and compiles C source code (lo_socket_shim.c) at runtime using gcc. The resulting shared library (lo_socket_shim.so) is then injected into the LibreOffice (soffice) process via the LD_PRELOAD environment variable to shim Unix socket system calls.
[COMMAND_EXECUTION]: Multiple scripts execute external system utilities. scripts/office/soffice.py executes gcc to compile the socket shim and soffice for document conversion. scripts/thumbnail.py executes pdftoppm (part of Poppler) to convert PDFs to JPEG thumbnails. scripts/office/validators/redlining.py executes git diff for text comparison.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted Office documents which serves as an ingestion point for potentially malicious data. 1. Ingestion points: scripts/thumbnail.py and scripts/office/validate.py read content from user-provided .pptx and .docx files. 2. Boundary markers: Absent; the skill processes document content directly without specific markers to ignore embedded instructions. 3. Capability inventory: Subprocess execution of soffice, pdftoppm, gcc, and git (in scripts/office/soffice.py, scripts/thumbnail.py, and scripts/office/validators/redlining.py), plus file system writes. 4. Sanitization: Present; the skill uses defusedxml and lxml to parse XML content, which provides protection against common XML-based attacks like XXE.

pptx