fine-tuning-expert

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

• Third-party content exposure detected (medium risk: 0.30). The workflow’s runtime ingests dataset text supplied via the user’s JSONL/path (e.g., python validate_dataset.py --input data.jsonl and then load_dataset(... data_files=...) and dataset.map(format_prompt)), so outsider-authored free text can be posted into whatever training dataset source the workflow loads.

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

• Potentially malicious external URL detected (high risk: 0.90). The code calls AutoModelForCausalLM.from_pretrained(...) with trust_remote_code=True and the example base model "meta-llama/Llama-3.1-8B", which will fetch and execute remote model code from that repo at runtime (meta-llama/Llama-3.1-8B).