hugging-face-model-trainer

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

• Third-party content exposure detected (high risk: 0.85). Outsider free text can enter the LLM context via load_dataset(...) fetching dataset examples from an arbitrary Hub dataset specified at runtime (e.g., user-provided --dataset/dataset_name), where the training script then tokenizes and feeds the dataset’s messages/text content into the model during training.

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

• Potentially malicious external URL detected (high risk: 1.00). The skill includes runtime fetching and execution of remote scripts (e.g., hf_jobs example using "https://huggingface.co/datasets/mcp-tools/skills/raw/main/dataset_inspector.py") and the GGUF conversion script clones and builds code from "https://github.com/ggerganov/llama.cpp.git", so external URLs are fetched at runtime and execute remote code.