The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes external system utilities including soffice (LibreOffice), git, and pdftoppm. These calls, located in ooxml/scripts/pack.py and ooxml/scripts/validation/redlining.py, are used for document integrity validation and text comparison. The implementation uses command arrays (lists) rather than shell strings, preventing shell injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: Setup instructions recommend installing dependencies from trusted system repositories and official package managers (apt, npm, and pip). These references target well-known technology providers and official registries.
[DATA_EXPOSURE]: The skill handles user-provided Office documents, which are essentially ZIP archives. The ooxml/scripts/unpack.py script includes explicit security checks to detect and reject unsafe archive members (symlinks and path traversal attempts). Furthermore, all XML processing is performed using the defusedxml library to mitigate XXE risks.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted document content which is eventually exposed to the AI agent's context as markdown or text. This represents a standard indirect prompt injection surface for analysis skills. The risk is managed by treating document content as data rather than executable instructions, alongside the platform's native safety guardrails.

docx-official