Towards Reliable LLM-Driven Fuzz Testing: Vision and Road Ahead