fine-tuning-expert

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

• Third-party content exposure detected (high risk: 0.70). The runtime training workflow ingests dataset text from data.jsonl / train.jsonl / test.jsonl via load_dataset(...) and dataset.map(format_prompt), so if those JSONL files contain outsider-authored free text, that text becomes LLM-readable context during fine-tuning (e.g., dataset_text_field="text" passed into SFTTrainer).

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

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly loads Hugging Face model repositories at runtime and in deployment code uses trust_remote_code=True (e.g., AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B")), which fetches and can execute remote repository code from https://huggingface.co/meta-llama/Llama-3.1-8B, so remote code is executed at runtime.