The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the subprocess module to execute system utilities for file conversion and comparison. \n
scripts/thumbnail.py invokes soffice and pdftoppm for rendering slide images. \n
scripts/office/validators/redlining.py calls git diff to compare text content between document versions. \n
scripts/office/soffice.py executes gcc to compile a library shim. \n- [REMOTE_CODE_EXECUTION]: The skill performs dynamic code generation and library injection to modify the behavior of external tools. \n
scripts/office/soffice.py compiles hardcoded C source code into a shared object file at runtime using gcc. \n
The library is then injected into the soffice process memory using the LD_PRELOAD environment variable to shim socket-related system calls. \n
While documented for environment compatibility, the ability to compile and inject code represents a high-privilege pattern. \n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection when processing document content. \n
Ingestion points: Untrusted data enters the agent's context through text and XML parsed from user-provided documents by scripts/office/unpack.py and the markitdown module. \n
Boundary markers: Absent. The skill does not employ delimiters or explicit instructions to delineate document content from agent instructions. \n
Capability inventory: The skill environment has access to shell command execution through the subprocess calls listed in the command execution finding. \n
Sanitization: While XML parsing is hardened using defusedxml, the extracted content itself is not sanitized for potential instruction-based payloads before being processed by the agent.

pptx