The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The helper script scripts/office/soffice.py performs dynamic execution by compiling a C shim library at runtime and injecting it into the LibreOffice process.
Evidence: The script contains a embedded C source string (_SHIM_SOURCE) that is written to a temporary file, compiled with gcc -shared -fPIC, and loaded via the LD_PRELOAD environment variable to intercept and modify system calls related to networking and socket operations.
[COMMAND_EXECUTION]: Several scripts utilize subprocess.run to execute system utilities and compilers:
scripts/office/soffice.py executes gcc to build the socket shim library.
scripts/thumbnail.py executes soffice (LibreOffice) to convert presentations to PDF and pdftoppm to generate slide thumbnails.
scripts/office/validators/redlining.py executes git diff to perform document comparisons during validation.
[PROMPT_INJECTION]: The skill exposes an indirect prompt injection surface by processing untrusted PowerPoint files.
Ingestion points: Data from external .pptx files is ingested via scripts/office/unpack.py and scripts/thumbnail.py.
Boundary markers: Absent. The scripts do not use delimiters or instructions to prevent the agent from following commands embedded in the slide content.
Capability inventory: The skill can execute shell commands, compile code, write to the file system, and utilize network-capable tools like LibreOffice.
Sanitization: The skill uses defusedxml to mitigate XML-specific attacks (like XXE), but does not sanitize extracted text content against adversarial AI instructions.
[EXTERNAL_DOWNLOADS]: The skill documentation and dependency list include several well-known third-party packages and system services:
Python packages: markitdown and Pillow.
Node.js packages: pptxgenjs, react-icons, react, react-dom, and sharp.
System dependencies: LibreOffice and Poppler.

pptx