跳到正文
arXiv cs.CL· Ziqi Miao, Lijun Li, Yuan Xiong, Zhenhua Liu, Pengyu Zhu, Chen Qian, Haonan Jia, Zihang Xu, Ken Deng, Jing Shao·· 9 小时前AI 评分40

Response Attack:利用上下文启动越狱大语言模型

Response Attack: Exploiting Contextual Priming to Jailbreak Large Language Models

AI 导读

研究者提出 Response Attack(RA)框架,利用对话中先前产生的中间、轻度有害回复作为上下文启动,在发出目标触发提示词前注入这些回复,从而诱导 LLM 输出违规内容。在八个 SOTA LLM 上的实验显示,RA 的攻击成功率显著高于九个主流越狱基线,且保持隐蔽性与效率。代码与数据已公开。

正文 · 原文

View PDF HTML (experimental)

Abstract:Contextual priming, where earlier stimuli covertly bias later judgments, offers an unexplored attack surface for large language models (LLMs). We uncover a contextual priming vulnerability in which the previous response in the dialogue can steer its subsequent behavior toward policy-violating content. While existing jailbreak attacks largely rely on single-turn or multi-turn prompt manipulations, or inject static in-context examples, these methods suffer from limited effectiveness, inefficiency, or semantic drift. We introduce Response Attack (RA), a novel framework that strategically leverages intermediate, mildly harmful responses as contextual primers within a dialogue. By reformulating harmful queries and injecting these intermediate responses before issuing a targeted trigger prompt, RA exploits a previously overlooked vulnerability in LLMs. Extensive experiments across eight state-of-the-art LLMs show that RA consistently achieves significantly higher attack success rates than nine leading jailbreak baselines. Our results demonstrate that the success of RA is directly attributable to the strategic use of intermediate responses, which induce models to generate more explicit and relevant harmful content while maintaining stealth, efficiency, and fidelity to the original query. The code and data are available at this https URL.
Comments: 20 pages, 10 figures. Code and data available at this https URL
Subjects: Computation and Language (cs.CL)
Cite as: arXiv:2507.05248 [cs.CL]
  (or arXiv:2507.05248v3 [cs.CL] for this version)
  https://doi.org/10.48550/arXiv.2507.05248

arXiv-issued DOI via DataCite

Journal reference: AAAI 2026

Submission history

From: Miao Ziqi [view email]
[v1] Mon, 7 Jul 2025 17:56:05 UTC (8,830 KB)
[v2] Fri, 21 Nov 2025 05:18:44 UTC (9,550 KB)
[v3] Tue, 17 Nov 2026 13:11:32 UTC (9,548 KB)

来源:arXiv cs.CL · arxiv.org