The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill communicates with well-known cloud services including the Baidu PaddleOCR-VL API for text recognition and OpenAI/Codex APIs for image generation and background repair. It downloads image results from provider-supplied URLs using a security-hardened client that includes SSRF protection (IP verification, host validation, and redirect prevention).
[COMMAND_EXECUTION]: The skill invokes local system utilities necessary for its primary function, including LaTeX engines (xelatex, lualatex, or pdflatex) for high-fidelity formula rendering and LibreOffice (soffice) or Microsoft PowerPoint for document conversion and visual QA. All subprocess calls use explicit paths or verified executables within the current environment.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input data (images, PDFs, and PPTX files) to extract content for reconstruction. This creates a surface for indirect prompt injection via OCR or metadata. The skill mitigates this risk by employing a structured workflow where extracted content is handled as advisory 'hints' and by confining sub-agent execution to a limited local directory scope without elevated privileges.
[SAFE]: The skill follows security best practices for secret management, storing API tokens in a dedicated configuration directory with restricted file permissions (chmod 600) and providing clear documentation regarding network boundaries and the handling of sensitive materials.

nature-image2ppt