The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/office/soffice.py performs runtime C code compilation and library injection. It writes source code to a temporary file /tmp/lo_socket_shim.c and uses gcc to compile it into a shared object library. This library is then injected into the soffice process via the LD_PRELOAD environment variable to intercept and redirect system socket calls.\n- [COMMAND_EXECUTION]: Several scripts utilize subprocess.run to execute external system binaries. Evidence includes scripts/office/soffice.py executing gcc and soffice, scripts/thumbnail.py executing pdftoppm, and scripts/office/validators/redlining.py executing git for diffing content.\n- [INDIRECT_PROMPT_INJECTION]: The skill extracts text from user-provided PowerPoint files, creating a vulnerability surface for indirect prompt injection.\n
Ingestion points: Uses markitdown to extract text content from PowerPoint files (.pptx, .potx) as described in SKILL.md.\n
Boundary markers: Slide number markers () are used to delimit content, but the skill lacks explicit instructions for the agent to ignore or isolate potentially malicious instructions embedded within that text.\n
Capability inventory: The skill possesses significant execution capabilities, including runtime code compilation (gcc), binary execution (subprocess.run), and file system manipulation (rezip, safe_extract).\n
Sanitization: The skill correctly utilizes defusedxml across its validation and cleaning scripts (scripts/clean.py, scripts/office/validate.py) to mitigate XML-based attacks like XXE.

pptx