The Agent Skills Directory

[PROMPT_INJECTION]: No malicious prompt injection patterns detected. The instructions are focused on guiding the agent to maintain academic rigor, source faithfulness, and proper citation discipline.
[DATA_EXFILTRATION]: No unauthorized data exfiltration patterns found. The skill performs legitimate network requests to fetch images from established stock providers (Wikimedia, Pexels, Pixabay) and generate audio via known cloud TTS services (ElevenLabs, MiniMax, Alibaba, etc.).
[CREDENTIALS_UNSAFE]: No hardcoded credentials detected. The skill uses standard environment variables and .env templates for managing API keys required by external AI providers.
[COMMAND_EXECUTION]: The skill uses subprocess.run to invoke trusted local tools like pandoc for document conversion and soffice (LibreOffice) for PDF generation. These operations are restricted to specific binaries with controlled arguments.
[EXTERNAL_DOWNLOADS]: The skill downloads project assets (images and audio) from reputable services. It also integrates with MinerU for PDF preprocessing, which is a recognized service in the academic space.
[DYNAMIC_EXECUTION]: Dynamic module loading is used in the image generation and search scripts to load provider backends. This is restricted by internal registries (BACKEND_REGISTRY, PROVIDER_MODULES), preventing the loading of arbitrary or untrusted modules.
[INDIRECT_PROMPT_INJECTION]: The skill has a vulnerability surface for indirect prompt injection because it processes untrusted user documents (PDF, DOCX, Web). However, the risk is inherent to its primary purpose, and the instructions to treat source material as evidence and prioritize source claims provide a mitigation against accidental obedience to embedded instructions.

academic-pptx