The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/office/soffice.py implements a process injection mechanism. It contains hardcoded C source code designed to shim system-level socket functions. At runtime, the script writes this source to disk, compiles it using gcc, and then loads the resulting shared library into the soffice (LibreOffice) process via the LD_PRELOAD environment variable.
[COMMAND_EXECUTION]: The skill uses subprocess.run to execute external binaries across multiple scripts:
scripts/office/soffice.py executes the gcc compiler to build the shared object for library injection.
scripts/accept_changes.py executes soffice with custom environment variables and macro injection to manipulate documents.
scripts/office/validators/redlining.py executes git diff to perform content comparison.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted document data as XML, creating an attack surface for indirect prompt injection.
Ingestion points: scripts/office/unpack.py reads user-provided Word documents and extracts their XML content.
Boundary markers: Absent; the skill does not use delimiters or instructions to ignore embedded commands in the processed XML.
Capability inventory: Extensive, including runtime compilation (gcc), document macro execution (soffice), and file system manipulation.
Sanitization: The skill uses defusedxml and lxml for parsing, which mitigates standard XML external entity (XXE) attacks, but the processed content still flows into sensitive logic paths.

docx