Research map / Large language models / Language, safety and interpretability
CHASE: Contextual History for Adaptive and Simple Exploitation in Large Language Model Jailbreaking
Zhiqiang Hao, Chuanyi Li, Ye Fan, Jun Cai, Xiao Fu, Shangqi Wang, Hao Shen, Jiao Yin, Jidong Ge, Bin Luo, Vincent Ng
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 |
|---|---|---|---|
| transfer jailbroken histories from an easy-to-jailbreak LLM to the target and prompt continuation (CHASE) instead of best baseline attacks | target LLMs · AdvBench jailbreak | ASR (%): 89.2 -> 95.8 (GPT-4, DRA vs CHASE) | Improved |
| CHASE versus DRA on Llama3-70b | Llama3-70b · AdvBench jailbreak | ASR (%): 85.2 -> 85.8 | No clear effect |
| CHASE requires 2-3 queries versus CodeAttack 20.80-30.31 | attack efficiency · AdvBench efficiency | average queries: 20.80 -> 2.12 (GPT-3.5) | Improved |
| CHASE versus LLM-Fuzzer on preparation cost | attack efficiency · AdvBench efficiency | QBC: 285.6 -> 8.5 | Improved |
| replace CHASE history generation with DRA to produce histories (ablation) | CHASE · AdvBench jailbreak | ASR (%): 98.3 -> 21.7 (GPT-3.5) | Hurt |
| replace CHASE history generation with CFA (ablation) | CHASE · AdvBench jailbreak | ASR (%): 98.3 -> 61.3 (GPT-3.5) | Hurt |
| increase number of history-generation turns from one to four (ablation) | history generation on Qwen1.5-72B · AdvBench history generation | ASR (%): 72.1 -> 100.0 | Improved |
| add LlamaGuard3-8B prompt filter defense (ablation) | CHASE · AdvBench jailbreak | ASR (%): 98.3 -> 78.8 (GPT-3.5) | Hurt |
| add refusal reminder to user prompt as defense (ablation) | CHASE · AdvBench jailbreak | ASR (%): 98.3 -> 15.4 (GPT-3.5) | Hurt |
| add refusal defenses (Guard, UsrPrompt, SysPrompt) to the target model | target LLMs · AlpacaEval | win rate (%): 51.3 -> 45.0 (GPT-4o, Guard) | Hurt |
Closest papers by meaning
- Multi-Turn Jailbreaking Large Language Models via Attention Shifting (AAAI 2025 · similarity 0.90)
- An Interpretable N-gram Perplexity Threat Model for Large Language Model Jailbreaks (ICML 2025 · similarity 0.90)
- Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models (NeurIPS 2026 · similarity 0.90)
- AutoDAN: Generating Stealthy Jailbreak Prompts on Aligned Large Language Models (ICLR 2024 · similarity 0.89)
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (NeurIPS 2024 · similarity 0.89)
- Many-shot Jailbreaking (NeurIPS 2024 · similarity 0.89)
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.
