huggingface-lora-space-builder

HIGH W007: Insecure credential handling detected in skill instructions.

• Insecure credential handling detected (high risk: 1.00). The skill explicitly instructs the agent to ask the user to "Paste it here" for a Hugging Face access token and then reuse that token in API calls (e.g. create/upload/add_space_secret, CLI predict examples), which requires the LLM to receive and potentially embed the secret verbatim in outputs or commands.

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

• Third-party content exposure detected (high risk: 0.75). Yes: in Phase 1 the workflow fetches outsider-authored LoRA model-card/README text from the Hub (e.g., ModelCard.load(repo_id) and card.text), which is then read/used to construct the agent’s LLM-visible context (trigger words, example prompts/media, and other prose-derived fields).

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

• Potentially malicious external URL detected (high risk: 0.80). The skill explicitly instructs loading remote code at runtime (e.g., client JS via the head script tag and packages installed from GitHub), such as https://cdnjs.cloudflare.com/ajax/libs/three.js/r128/three.min.js and git+https://github.com/huggingface/diffusers (and git+https://github.com/Lightricks/LTX-Video.git), which will be fetched and executed as part of the Space runtime/build.