huggingface-llm-trainer

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

• Third-party content exposure detected (medium risk: 0.30). The skill’s runtime training/validation workflow ingests outsider-authored free text from user-supplied Hugging Face datasets via datasets.load_dataset()/load_dataset(...).map(...) (e.g., dataset_split in SKILL.md and dataset formatting in scripts/train_sft_example.py / references/local_training_macos.md), so attacker-published dataset rows can be read by the LLM during prompt construction/training.

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

• Potentially malicious external URL detected (high risk: 0.90). High-confidence flag: the skill fetches and executes remote code at runtime (e.g., cloning/executing llama.cpp via "https://github.com/ggerganov/llama.cpp.git" and running raw script URLs like "https://raw.githubusercontent.com/huggingface/trl/main/examples/scripts/grpo.py" and Hugging Face raw script URLs for the dataset inspector), which are runtime external dependencies that execute remote code.