This paper presents a novel framework that consolidates multi-turn adversarial "jailbreak" prompts into single-turn queries, significantly reducing the manual effort required for adversarial testing of large-scale language models (LLMs). Multi-turn human jailbreaks have shown high attack success rates but require significant human resources and time. The proposed multi-turn-single-turn (M2S) method (Hyphenize, Numberize, Pythonize) systematically reformats multi-turn conversations into structured single-turn prompts. Despite eliminating repetitive interactions, these prompts maintain and often improve adversarial efficacy. In extensive evaluations on the Multi-turn Human Jailbreak (MHJ) dataset, the M2S method achieves attack success rates ranging from 70.6% to 95.9% on several state-of-the-art LLMs. Remarkably, the single-turn prompts outperform the original multi-turn attack by up to 17.5 percentage points and reduce average token usage by more than half. Further analysis reveals that embedding malicious requests in structures like enumerations or codes exploits "contextual blind spots" to bypass both basic safeguards and external input/output filters. The M2S framework transforms multi-round conversations into concise, single-round prompts, providing a scalable tool for large-scale adversarial testing and exposing a critical weakness in modern LLM defenses.
