The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/office/soffice.py dynamically generates and executes code to facilitate LibreOffice operations in restricted environments.
It contains a C source code string (_SHIM_SOURCE) designed to shim system-level socket calls.
The script executes gcc at runtime to compile this source into a shared object file (lo_socket_shim.so) in the system's temporary directory.
It utilizes the LD_PRELOAD environment variable to inject the compiled library into the soffice process execution context.
[COMMAND_EXECUTION]: Several scripts within the skill utilize the subprocess module to execute system commands for document processing and maintenance.
scripts/office/soffice.py invokes gcc for compilation and soffice for document conversion.
scripts/thumbnail.py executes pdftoppm (part of the Poppler suite) to generate image previews from PDF versions of slides.
scripts/office/validators/redlining.py uses git diff to compare text content and identify edits that have not been properly tracked.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PowerPoint files which are parsed as structured XML, creating a potential attack surface for indirect injection.
Ingestion points: External content from .pptx or .potx files is loaded and parsed in scripts/add_slide.py, scripts/clean.py, and the validation scripts.
Boundary markers: The skill does not employ clear delimiters or instructions to prevent the agent from inadvertently executing commands or following instructions embedded within slide content.
Capability inventory: The skill environment permits shell command execution and runtime code compilation, which could be abused if an agent follows instructions found in malicious files.
Sanitization: While the skill correctly uses defusedxml to mitigate standard XML-based exploits, the extracted text content is returned to the agent context without additional sanitization.

pptx