The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill performs dynamic code generation and process injection. The script scripts/office/soffice.py writes embedded C source code to the filesystem, compiles it at runtime using gcc, and then loads the resulting library into the soffice process using LD_PRELOAD to hook system calls. Additionally, scripts/accept_changes.py dynamically generates StarBasic macros to automate document modifications through LibreOffice.\n- [COMMAND_EXECUTION]: Several scripts within the skill utilize the subprocess module to execute system-level utilities. These include the invocation of gcc for compilation, git for comparing document content, and soffice for document format conversion and processing.\n- [EXTERNAL_DOWNLOADS]: The documentation in SKILL.md directs the user to perform an unversioned global installation of the docx library from the npm registry (npm install -g docx), which introduces an external dependency and supply chain risk at the time of skill configuration.\n- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to its ingestion of untrusted external document data and its high-privilege system capabilities.\n
Ingestion points: The scripts scripts/office/unpack.py and scripts/office/pack.py process data from ZIP archives and constituent XML files within .docx documents.\n
Boundary markers: No explicit markers or delimiters are used to isolate untrusted document content from the agent's control instructions.\n
Capability inventory: The skill possesses extensive system access through shell execution, compiler invocation, and filesystem operations across multiple scripts.\n
Sanitization: While the skill uses defusedxml to mitigate some XML-based attacks, it does not sanitize document content before it influences the behavior of system-level utilities.

docx