The paper evaluates six automated jailbreak evaluators—HarmBench, JailbreakBench, JailbreakRadar, StrongReject, JADES, and JailMeter—using human-labeled data from JailbreakQR and JailMeter-Eva. It measures each evaluator’s agreement with human judgments, error types, and consistency across attack families, controlling for model-specific variation by using a shared LLM judge where needed. The study finds that JADES performs best overall, with HarmBench and StrongReject also showing strong performance.
By Yujie Mu
arXiv:2607. 19424v1 Announce Type: cross Abstract: The assessment of jailbreak attacks against large language models currently suffers from inconsistent evaluation criteria and methods, leading to unreliable estimates of attack success rates.
By Qingjia Huang, Jingyu Zhang, Jianguo Wu, Yakai Li, Weijuan Zhang, Yankai Rong, Junyi Yao, Shengzhi Zhang, Xiaoqi Jia
The paper introduces Fair-ASR, a protocol for evaluating black‑box jailbreak attacks using a shared target‑call budget, addressing the bias of prior studies that rely solely on attack success rate. Re‑evaluating 11 attacks under this protocol shows significant shifts in rankings and highlights that many methods are not efficient in both target and attacker calls. The authors then present ReCode, a budget‑efficient attack that combines desensitization rewriting with low‑cost primitives, achieving 85% ASR on GPT‑5 with only 7.19 attacker calls per request under a 20‑target‑call budget.
By Zhida He, Xiaoyu Wen, Han Qi, Ziyuan Zhou, Peng Yu, Jiajia Li, Chaochao Lu, Qiaosheng Zhang
arXiv:2606. 11425v1 Announce Type: cross Abstract: Jailbreak attacks expose persistent safety weaknesses in large language models (LLMs), but existing stateless single-turn methods face a trade-off: hand-crafted prompts are expressive but static, while iterative prompt optimization can adapt but often relies on low-level mutations that require many target queries.
By Ge Shi, Jun Yin, Donglin Xie, Fangyi Liu, Yucan Li, Menglin Liu
arXiv:2602. 24009v4 Announce Type: replace-cross Abstract: Jailbreak techniques for large language models (LLMs) evolve faster than benchmarks, making robustness estimates stale and difficult to compare across papers due to drift in datasets, harnesses, and judging protocols.
By Zhicheng Fang, Jingjie Zheng, Chenxu Fu, Wei Xu
The paper introduces Fair-ASR, a new evaluation protocol for black-box jailbreak attacks that uses shared target-call budgets to provide a fair comparison across methods. Re‑evaluating 11 attacks under this protocol shows that rankings shift significantly with different budgets, and that simple perturbations and templates remain competitive. The authors also present ReCode, a budget‑efficient attack that combines desensitization rewriting with low‑cost primitives, achieving 85% ASR on GPT‑5 with only 7.19 attacker calls per request under a 20‑call budget.
The paper introduces TACS, a trajectory‑aware candidate selection framework designed to improve jailbreak suffix optimization for large language models. Traditional gradient‑based methods choose candidates based solely on the lowest current loss, which the authors argue is myopic and can lead to reward hacking. TACS augments per‑step evaluation with a trajectory‑aware proxy, reference‑policy regularization, and a chi‑squared correction to encourage selections that remain effective beyond the immediate step. Experiments on HarmBench show that TACS consistently outperforms strong baselines, achieving higher attack success rates and more stable optimization behavior.
By Shiliang Xiao
AlcaTRAz is a prompt‑level defense that uses rule trees to insert controlled character‑level perturbations into input text, disrupting jailbreak attacks without modifying or retraining the target LLM. It operates solely on the input, making it suitable for black‑box deployments, and was evaluated on 33 open‑weight models and 22 jailbreak types, outperforming three baseline defenses in 73.4 % of model‑attack combinations. While it significantly reduces high‑severity jailbreak success, it does not eliminate it and is intended as one layer of a broader defense strategy.
By Jakub Re\v{s}, Petr Ka\v{s}ka, Martin Pere\v{s}\'ini, Martin Ukrop, Kamil Malinka
The paper presents a systematic study of combining defenses against jailbreak attacks on Large Language Models across different pipeline stages. It introduces a standardized evaluation framework that defines attack-success-rate, controls query budgets, and applies explicit fairness rules. Across 19 attacks and 15 defenses, the study finds that no single defense dominates, but carefully selected combinations can provide strong safety with minimal loss of utility, offering practical guidance for layered defense pipelines.
By Jiale Luo, Eric Han
The paper "Jailbreaking in the Haystack" introduces NINJA, a jailbreak technique that exploits long-context language models by appending benign, model-generated content to harmful user goals. It demonstrates that the position of harmful goals within the context is crucial for safety, and shows that NINJA significantly boosts attack success rates on models such as LLaMA, Qwen, Mistral, and Gemini. Unlike previous methods, NINJA is low-resource, transferable, less detectable, and compute‑optimal, revealing that carefully crafted benign long contexts can expose fundamental vulnerabilities in modern LMs.
By Rishi Rajesh Shah, Chen Henry Wu, Shashwat Saxena, Ziqian Zhong, Alexander Robey, Aditi Raghunathan
The paper investigates how much of a multi-turn jailbreak attack’s query history is necessary to achieve the same progress. By removing early attacker queries and re-running the remaining sequence, the authors define the effective horizon as the shortest query set that reproduces the attack outcome. They find that, on average, about half of the original queries suffice, a phenomenon they term effective-horizon compression, and demonstrate that this shorter history can improve attack continuation, training efficiency, and maintain defense calibration.
By Zhipeng Wei, Michael Lee, N. Benjamin Erichson
The study evaluates how identically trained language models respond to jailbreak prompts, finding that models trained with the same pretraining run exhibit highly correlated non-refusal rates (average 0.984), while different pretraining seeds reduce correlation to 0.811 and different data recipes slightly lower it further to 0.779. This indicates that a model’s susceptibility to jailbreaks is largely determined by its specific pretraining run rather than the data used. Additionally, each pretraining run leaves a distinct fingerprint, enabling identification of the run from a model’s responses to a set of prompts.
By Chi-Wei Chang, Mu-Tien Kuo, Kuan-Wei Lee, Meng-Hsun Tsai