The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/office/soffice.py performs runtime compilation of a hardcoded C source file into a shared object (lo_socket_shim.so) using gcc. This library is then injected into the soffice process environment via the LD_PRELOAD variable. While intended as a workaround for socket restrictions in sandboxed environments, this represents a sophisticated process injection and dynamic execution pattern.
[COMMAND_EXECUTION]: Several scripts utilize subprocess.run to execute external system utilities. For instance, scripts/thumbnail.py invokes soffice and pdftoppm for slide rendering, while scripts/office/validators/redlining.py executes git diff to compare document revisions.
[PROMPT_INJECTION]: The skill's primary function is to process untrusted external PowerPoint files, which constitutes an indirect prompt injection surface.
Ingestion points: Data enters the agent's context through text extraction from .pptx files as directed in SKILL.md and editing.md.
Boundary markers: The instructions lack explicit delimiters or warnings to ignore instructions embedded within the processed presentation content.
Capability inventory: The skill leverages powerful system capabilities, including a compiler (gcc), document converters (soffice), and file system access, which could be leveraged if malicious content is processed.
Sanitization: There is no evidence of sanitization, filtering, or validation for the natural language content extracted from user-provided presentation files.

alterlab-pptx