fine-tuning-expert

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

• Third-party content exposure detected (high risk: 0.75). The skill’s runtime workflow can ingest outsider-authored free text via load_custom_dataset(raw_data_path) which calls load_dataset(path, split="train") for arbitrary Hugging Face Hub dataset identifiers (public web content), and then passes the resulting instruction/input/output/messages strings into LLM-relevant training/evaluation code.

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

• Potentially malicious external URL detected (high risk: 1.00). The skill fetches and loads remote model repositories at runtime (e.g., AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B", trust_remote_code=True)) and references a runtime container image (ghcr.io/huggingface/text-generation-inference:latest), which cause remote code and artifacts to be downloaded and potentially executed, creating a clear runtime external code-execution dependency.