I suspect that distillation attacks may be slightly exaggerated. Most of the training data used during fine-tuning is now synthetic data. You can't just repeat the same stuff twice, therefore another LLM is writing a text book that is explaining a topic in detail, ideally without any gaps in the material.