Prefill-level Jailbreak: A Black-Box Risk Analysis of Large Language Models