The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/office/soffice.py implements a runtime compilation workflow where C source code is generated and compiled using gcc into a shared library. This library is then injected into the soffice process via the LD_PRELOAD environment variable to modify system-level socket operations. This matches patterns of sophisticated dynamic code execution and process manipulation.
[COMMAND_EXECUTION]: The skill relies on spawning multiple subprocesses to invoke system utilities. Evidence includes calls to gcc in scripts/office/soffice.py, calls to soffice and pdftoppm in scripts/thumbnail.py, and calls to git diff in scripts/office/validators/redlining.py.
[PROMPT_INJECTION]: The skill provides an indirect prompt injection surface by extracting text from untrusted PPTX files using markitdown. Ingestion point: Slide text extracted via python -m markitdown as instructed in SKILL.md. Boundary markers: Absent; extracted content is directly presented to the agent context. Capability inventory: Shell command execution via subprocess.run (thumbnail.py, soffice.py, redlining.py) and file system access. Sanitization: No filtering or escaping of extracted text is implemented. This creates a risk where malicious instructions in a document could influence the AI agent's behavior.

pptx