The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Runtime C code compilation and process injection.\n
File: scripts/office/soffice.py\n
The skill hardcodes a C source string (_SHIM_SOURCE) and compiles it into a shared object library (lo_socket_shim.so) using gcc at runtime.\n
This compiled library is then injected into the LibreOffice (soffice) process environment via the LD_PRELOAD environment variable to shim networking behavior in restricted environments.\n- [COMMAND_EXECUTION]: Execution of system binaries and runtime macro generation.\n
File: scripts/recalc.py\n
The skill dynamically generates a LibreOffice Basic macro (Module1.xba) and writes it to a temporary profile directory to automate formula recalculation operations within the spreadsheet.\n
Files: scripts/recalc.py, scripts/office/soffice.py, scripts/office/validators/redlining.py\n
The skill invokes several external tools via subprocess.run, including soffice (LibreOffice), gcc (GNU Compiler Collection), git, and system timeout utilities (timeout, gtimeout).\n- [PROMPT_INJECTION]: Indirect prompt injection attack surface.\n
Files: SKILL.md, scripts/recalc.py\n
Ingestion points: The skill ingests untrusted data from user-provided Office documents (XLSX, XLSM, DOCX, PPTX) which are then processed by various scripts and tools.\n
Boundary markers: There are no explicit delimiters or instruction markers used when extracting and processing data from these spreadsheets, meaning malicious instructions in a cell could potentially influence the agent.\n
Capability inventory: The skill's ability to execute system commands, compile code, and perform process injection represents a powerful capability set that increases the impact of a successful indirect injection attack.\n
Sanitization: No sanitization or filtering is performed on the data content extracted from the spreadsheets before it is passed to internal automation logic.

xlsx