Research map / Large language models / Language, safety and interpretability
Response Attack: Exploiting Contextual Priming to Jailbreak Large Language Models
Miao Ziqi, Lijun Li, Yuan Xiong, Zhenhua Liu, Pengyu Zhu, Jing Shao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add injected mildly harmful response and trigger prompt (Rharm) (ablation) | RA-DRI · HarmBench (avg over 8 LLMs) | ASR (%): 48.5 -> 78.8 | Improved |
| add prompt rewriting of query into Pinit (ablation) | RA-DRI · HarmBench (avg over 8 LLMs) | ASR (%): 57.2 -> 78.8 | Improved |
| add prompt rewriting of query into Pinit (ablation) | RA-SRI · HarmBench (avg over 8 LLMs) | ASR (%): 29.9 -> 73.3 | Improved |
| add prompt rewriting on GPT-4.1 (ablation) | RA-SRI · HarmBench (GPT-4.1) | ASR (%): 4.5 -> 71.0 | Improved |
| include initial user query Pinit in priming dialogue (ablation) | RA-DRI · HarmBench (avg over 8 LLMs) | ASR (%): 75.6 -> 78.7 | No clear effect |
| generate response from rewritten prompt rather than original query (ablation) | RA-DRI · HarmBench (avg over 8 LLMs) | ASR (%): 60.6 -> 75.6 | Improved |
| use a semantically rich injected response instead of affirmative prefix (ablation) | RA-DRI · HarmBench (avg over 8 LLMs) | ASR (%): 40.5 -> 78.7 | Improved |
| multi-turn chat formatting vs single-turn flattened formats (ablation) | RA-DRI · HarmBench (avg over 8 LLMs) | ASR (%): 78.7 -> 74.6/74.8 | No clear effect |
| RA-DRI vs strongest baseline ActorAttack | RA-DRI · HarmBench (avg over 8 LLMs) | ASR (%): 82.8 -> 94.8 | Improved |
| RA-SRI vs FlipAttack on LLaMA-3-8B | RA-SRI · HarmBench (LLaMA-3-8B) | ASR (%): 0.0 -> 76.0 | Improved |
| RA-SRI vs FlipAttack on GPT-4.1 | RA-SRI · HarmBench (GPT-4.1) | ASR (%): 89.5 -> 88.0 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FlipAttack — FlipAttack: Jailbreak LLMs via Flipping (ICML 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Many-shot Jailbreaking (NeurIPS 2024 · similarity 0.88)
- Multi-Turn Jailbreaking Large Language Models via Attention Shifting (AAAI 2025 · similarity 0.88)
- Measuring Non-Adversarial Reproduction of Training Data in Large Language Models (ICLR 2025 · similarity 0.87)
- REINFORCE Adversarial Attacks on Large Language Models: An Adaptive, Distributional, and Semantic Objective (ICML 2025 · similarity 0.87)
- Adjacent Words, Divergent Intents: Jailbreaking Large Language Models via Task Concurrency (NeurIPS 2025 · similarity 0.87)
- Cannot See the Forest for the Trees: Invoking Heuristics and Biases to Elicit Irrational Choices of LLMs (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
