The Agent Skills Directory

[SAFE]: The skill implements an exemplary security model focused on defense-in-depth. It uses a custom 'package admission' system in lib/package_safety.py that sanitizes DOCX files by rejecting macros, OLE objects, ActiveX controls, and external relationships before they are processed by higher-level parsers or renderers.
[COMMAND_EXECUTION]: The skill uses subprocess.run in scripts/render_docx.py and scripts/check_runtime.py to invoke system utilities like libreoffice and pdftoppm. These calls are restricted to purpose-specific tasks, and the skill attempts to use unshare to isolate these processes in a no-network namespace to prevent data exfiltration.
[PROMPT_INJECTION]: The skill addresses indirect prompt injection (Category 8) by providing explicit instructions in SKILL.md and references/safety.md to treat all document content as untrusted data rather than instructions. It includes a mandatory visual QA workflow where the agent must inspect rendered images of the document to verify layout and content integrity.
[SAFE]: File system operations are protected by lib/path_safety.py, which enforces atomic writes, prohibits path traversal via parent directory escapes, and explicitly rejects symbolic link components in both input and output paths.
[SAFE]: Dynamic imports using importlib.import_module in scripts/check_runtime.py are limited to a static, hardcoded list of trusted dependencies for environment validation.

krt-word-illuminator