The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill performs dynamic code generation and execution. Specifically, scripts/office/soffice.py contains hardcoded C source code that is written to a temporary file and compiled with gcc into a shared library. This library is then injected into the soffice process using the LD_PRELOAD environment variable to hook socket-related system calls. Additionally, scripts/accept_changes.py generates and writes a StarBasic macro to the LibreOffice application profile to facilitate automated document editing tasks.
[COMMAND_EXECUTION]: Several scripts, including scripts/office/soffice.py, scripts/accept_changes.py, and scripts/office/validators/redlining.py, utilize the subprocess module to invoke external binaries such as soffice, pandoc, pdftoppm, and the gcc compiler. This grants the skill the ability to execute arbitrary system commands and compilers on the host system.
[EXTERNAL_DOWNLOADS]: The skill documentation specifies the installation of the docx library from the global Node package manager registry (npm install -g docx). This is an expected external dependency for the skill's document generation functionality.
[PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection due to its processing of untrusted content from external Word documents.
Ingestion points: Document text and XML content are extracted and processed via pandoc and the scripts/office/unpack.py script.
Boundary markers: The instructions lack delimiters or explicit warnings to the agent to disregard instructions embedded within the ingested document data.
Capability inventory: The skill can execute shell commands, compile and run C code, and perform file system operations.
Sanitization: While the skill uses defusedxml for some XML operations, lxml is also employed for parsing and validation without explicit security hardening, potentially exposing the agent to XML-based vulnerabilities if processing malicious files.

docx