The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/office/soffice.py implements runtime C compilation and library injection. It contains a hardcoded C source string that is written to a temporary file and compiled into a shared object using gcc. This library is then preloaded into the soffice (LibreOffice) process via the LD_PRELOAD environment variable to shim network socket system calls. This is used to ensure compatibility with restricted sandbox environments but employs high-privilege execution techniques.
[COMMAND_EXECUTION]: The skill makes extensive use of the subprocess module to execute external binaries, including gcc for runtime compilation, soffice for document conversion, pdftoppm for image processing, and git for document diffing. While these operations appear legitimate for the skill's purpose, they provide a broad capability surface for the agent.
[INDIRECT_PROMPT_INJECTION]: The skill extracts text and structured data from user-provided PowerPoint (.pptx and .potx) files, which are processed by the agent. This creates a surface for indirect prompt injection if the source files contain hidden instructions.
Ingestion points: Content is extraction from presentation XML files using markitdown and custom scripts within the scripts/ directory.
Boundary markers: The extracted content is not encapsulated with specific instructions or delimiters to isolate it from the agent's operational logic.
Capability inventory: The skill possesses significant capabilities, including the ability to read and write files, execute subprocesses, and perform runtime compilation and library injection.
Sanitization: The skill mitigates common XML-based attacks (such as XXE) by consistently utilizing the defusedxml library for all XML parsing operations.

pptx