A preprint finds LLM-enhanced graph models score higher on classification while showing more link, label and membership leakage.