obsidian-vault-builder

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.80). The SKILL.md explicitly instructs Smart Connections / transformers.js to download ONNX models from public hosts (e.g., huggingface.co / Xenova paths described in the "Smart Connections language coverage" section) and references public services/plugins (Kroki, GitHub plugin releases) whose arbitrary third‑party model/artifact content would be ingested and used to produce embeddings/search results that can materially influence agent behavior.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.80). The skill explicitly states the plugin's transformers.js runtime will fetch model binaries from huggingface.co (e.g. huggingface.co//resolve/main/onnx/model_quantized.onnx), which are downloaded at runtime and executed as ONNX models to drive embeddings (thus executing external code that directly affects agent behavior).