fine-tuning-customization

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

• Third-party content exposure detected (high risk: 0.80). The skill's training scripts and examples (e.g., load_preference_data in scripts/dpo-training.py and multiple examples calling load_dataset("your_dataset") / load_dataset(..., split="train")) explicitly load datasets from Hugging Face or arbitrary dataset identifiers, which are public, user-contributed sources that the agent ingests as training/preference data and could include untrusted user-generated content.