The Agent Skills Directory

[COMMAND_EXECUTION]: Runtime compilation of a C shim library.
Evidence: In scripts/office/soffice.py, the skill includes a hardcoded C source string (_SHIM_SOURCE) that it writes to a file and compiles at runtime using gcc. The resulting shared object is stored in the user's cache directory.
[COMMAND_EXECUTION]: Process injection using LD_PRELOAD.
Evidence: In scripts/office/soffice.py, the skill executes the soffice binary with the LD_PRELOAD environment variable set to the path of the dynamically compiled shim. This is used to intercept and modify system calls related to socket operations, bypassing security restrictions in certain environments.
[COMMAND_EXECUTION]: Arbitrary execution of system binaries and macros.
Evidence: The skill invokes soffice, gcc, and git using subprocess.run. Furthermore, scripts/recalc.py configures and executes a LibreOffice Basic macro (RecalculateAndSave) by writing it to the application's configuration path and invoking it via command-line arguments.
[PROMPT_INJECTION]: Deceptive and inconsistent skill metadata.
Evidence: The SKILL.md metadata lists the author as "K-Dense Inc.", whereas the platform identifies the author as "hkphysics". Additionally, the description explicitly instructs the agent NOT to use the skill for Word or PowerPoint documents, yet the skill includes a complete suite of validation and processing logic (over 40 files including XSD schemas) specifically for .docx and .pptx formats.
[PROMPT_INJECTION]: Surface for indirect prompt injection via data ingestion.
Ingestion points: SKILL.md provides instructions and patterns for reading untrusted spreadsheet data directly into the agent's context using pandas and openpyxl.
Boundary markers: Absent; there are no instructions or delimiters defined to help the agent distinguish between data content and potential malicious instructions within the spreadsheets.
Capability inventory: The skill has extensive command execution capabilities through the Bash tool and specialized scripts like scripts/office/soffice.py.
Sanitization: Absent; content read from spreadsheets is not filtered or sanitized before being processed by the agent or by LibreOffice macros.

xlsx