The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/office/soffice.py performs runtime compilation of a C shim using gcc. The resulting shared library is then loaded into the LibreOffice process via LD_PRELOAD to wrap socket function calls. This represents a high-privilege capability involving both code generation and injection techniques.
[COMMAND_EXECUTION]: Multiple components execute system utilities via subprocess.run. This includes running gcc for shim compilation, git for document difference analysis in scripts/office/validators/redlining.py, and soffice and pdftoppm for slide image conversion in scripts/thumbnail.py.
[PROMPT_INJECTION]: The skill exhibits a significant surface for indirect prompt injection through its ingestion of external presentation data.
Ingestion points: Content is extracted from user-provided .pptx files using markitdown and visualized through image thumbnails generated by thumbnail.py.
Boundary markers: Analysis of SKILL.md and associated documentation reveals an absence of delimiters or instructions for the agent to ignore potentially adversarial commands embedded in extracted content.
Capability inventory: The skill possesses extensive local capabilities, including file system modification and execution of multiple system binaries, which could be leveraged if an agent obeys instructions hidden within a processed presentation.
Sanitization: There is no evidence of content filtering or sanitization performed on the extracted slide data before it is presented to the agent's context.

pptx