Researchers describe Response Attack, a jailbreak method using earlier dialogue replies as primers
An arXiv paper reports that a model's own previous response can steer it toward policy-violating content, with reported success across eight LLMs.
In plain words
A jailbreak is a way to trick an AI chatbot into breaking its safety rules. The paper's authors say that placing mildly harmful replies earlier in a conversation can prime the model, so a later trigger prompt makes it produce more harmful content.
What was announced
The authors frame contextual priming, in which earlier stimuli covertly bias later judgments, as an unexplored attack surface for large language models. Their framework, Response Attack, reformulates harmful queries and injects intermediate, mildly harmful responses into a dialogue before issuing a targeted trigger prompt. They position this against existing jailbreak approaches that rely on single-turn or multi-turn prompt manipulation or static in-context examples, which they say suffer from limited effectiveness, inefficiency, or semantic drift.
In experiments across eight state-of-the-art LLMs, the authors report that Response Attack achieved significantly higher attack success rates than nine leading jailbreak baselines. They attribute this to the strategic use of intermediate responses, which they say induced models to generate more explicit and relevant harmful content while maintaining stealth, efficiency, and fidelity to the original query. Code and data are posted in a public repository.
What it does not show
- The evidence is the paper's abstract only; results are author-reported and not independently confirmed, and the specific models, baselines, and success-rate figures are not given in the excerpt.
For security and IT teams
- Scope
- The described vulnerability concerns multi-turn dialogue behavior in LLMs, where earlier responses in the same conversation influence later output.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.