The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/office/soffice.py contains logic to compile a C source string (_SHIM_SOURCE) into a shared object file (lo_socket_shim.so) at runtime using gcc. This code is then used for process injection.
[DYNAMIC_EXECUTION]: The skill uses the LD_PRELOAD environment variable in scripts/office/soffice.py to inject the dynamically compiled shared library into the soffice (LibreOffice) process. This technique is used to intercept and redirect standard system calls related to socket communication.
[COMMAND_EXECUTION]: The skill executes multiple external command-line tools via subprocess calls, including soffice (LibreOffice), pandoc, pdftoppm, and gcc. These tools are necessary for document conversion, text extraction, and the compilation of the functional shim.
[INDIRECT_PROMPT_INJECTION]: The skill extracts text and structure from external .docx files, which provides a surface for indirect prompt injection if the documents contain adversarial content.
Ingestion points: Untrusted data enters the agent context through pandoc text extraction and XML file processing in scripts/office/unpack.py.
Boundary markers: The instructions do not specify the use of clear delimiters or instructions to ignore embedded commands when the agent reads extracted text.
Capability inventory: The skill has significant capabilities including executing system commands, compiling and injecting code, and performing complex file system operations.
Sanitization: The skill effectively mitigates XML External Entity (XXE) attacks by using the defusedxml library for all XML parsing operations.

docx