Jailbreaking Commercial Black-Box LLMs with Explicitly Harmful Prompts

TL;DR

This paper presents a novel attack method called DH-CoT that effectively jailbreaks state-of-the-art black-box LLMs by integrating multiple tricks into a single template, and introduces a data cleaning framework to evaluate attack success.

Contribution

It introduces DH-CoT, a new multi-trick jailbreak attack, and MDH, a data cleaning framework, to improve attack effectiveness and evaluation accuracy on recent LLMs.

Findings

DH-CoT outperforms SOTA methods like H-CoT and TAP.

MDH reliably filters low-quality samples for evaluation.

DH-CoT successfully jailbreaks models including GPT-5 and Claude-4.

Abstract

Existing black-box jailbreak attacks achieve certain success on non-reasoning models but degrade significantly on recent SOTA reasoning models. To improve attack ability, inspired by adversarial aggregation strategies, we integrate multiple jailbreak tricks into a single developer template. Especially, we apply Adversarial Context Alignment to purge semantic inconsistencies and use NTP (a type of harmful prompt) -based few-shot examples to guide malicious outputs, lastly forming DH-CoT attack with a fake chain of thought. In experiments, we further observe that existing red-teaming datasets include samples unsuitable for evaluating attack gains, such as BPs, NHPs, and NTPs. Such data hinders accurate evaluation of true attack effect lifts. To address this, we introduce MDH, a Malicious content Detection framework integrating LLM-based annotation with Human assistance, with which we clean data and build RTA dataset suite. Experiments show that MDH reliably filters low-quality samples and that DH-CoT effectively jailbreaks models including GPT-5 and Claude-4, notably outperforming SOTA methods like H-CoT and TAP.