REVIEW 4 major objections 6 minor 68 references
SoK: A Comprehensive Security Analysis of Jailbreak Resilience in GPT and DeepSeek Models
T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read The paper claims DeepSeek's jailbreak resistance is selective—it withstands optimization-driven attacks but yields much more than GPT-4 to prompt-based and human-crafted adversarial prompts.
desk verdict A useful DeepSeek-vs-GPT jailbreak benchmark whose central conclusion is an artifact of comparing raw ASR; under the paper's own DR-normalized metric, the susceptibility result reverses. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the attack success rate (ASR) computed by a fixed evaluation protocol: a curated set of 510 harmful behaviors, seven attack methods spanning black-box and white-box categories, and a fine-tuned classifier that labels each model completion as a successful jailbreak or a refusal. Behavior-level breakdowns by functional category (standard, copyright, contextual, multimodal) and by six high-risk semantic domains let the authors separate where safety holds from where it fails. The direct-request baseline serves as the reference point, and the paper argues that reporting the margin over this baseline captures scaling-related risk better than the raw ASR.
What would settle it
Re-score a stratified sample of completions—balanced by model family, attack method, and behavior category—with independent human raters using the paper's stated success criteria, then recompute the attack success rates; if the DeepSeek-over-GPT gap on prompt-based attacks shrinks below statistical significance, the central trade-off claim fails.
Extended reading notes
Core claim
The discovery is a systematic asymmetry in jailbreak resilience. Under the same evaluation protocol, DeepSeek models show partial resistance to the tree-search attack TAP-T, but their attack success rates under prompt-based attacks are much higher than GPT-4's: DeepSeek reaches roughly 77% on standard behaviors and 74% on contextual behaviors under a zero-shot prompting attack, about double GPT-4's rates, while GPT-4 Turbo stays lower and more uniform across categories. Within the DeepSeek family, gradient-based and chain-of-thought attacks become more successful as parameters grow, with GCG rising from 33.44% at 1.5B to 55.31% at 32B. The paper explains the gap by alignment strategy: the GPT family benefits from reinforcement-learning-from-human-feedback pipelines and red-teaming, while DeepSeek's distilled variants rely more on supervised fine-tuning. It concludes that scaling capability without scaling alignment enlarges the attack surface.
Load-bearing premise
The load-bearing premise is that the fine-tuned classifier correctly identifies jailbreak success for both model families; if that judge systematically favors or penalizes one family's style of completions, the reported attack-success-rate gaps are not trustworthy.
Editorial extensions
If this is right
- Larger DeepSeek variants (7B to 32B) are more, not less, vulnerable to gradient-based and chain-of-thought attacks, so model scaling without alignment scaling widens the attack surface.
- Defenses should prioritize gradient-based adversarial training and refusal mechanisms that monitor chain-of-thought reasoning, and evaluations should report ASR relative to the direct-request baseline.
- The safety gap is category-specific: DeepSeek is stronger on chemical/biological and copyright behaviors, while GPT models are stronger on misinformation, cybercrime, harassment, and illegal content.
- Open-weight releases need alignment pipelines with the same breadth of feedback and red-teaming as the GPT family, or deployment carries a measurable safety penalty.
Reading between the lines
- Inference: The failure pattern suggests a testable hypothesis—fine-tuning a DeepSeek variant on adversarial paraphrases from prompt-rewriting and persuasion attacks should close most of the zero-shot gap, because the vulnerability appears to be surface-form generalization rather than missing safety knowledge.
- Inference: Because the whole comparison depends on one classifier judge, a human-labeled replication on a few hundred outputs would settle whether the family-level gaps are real; the direction of the comparison could shift if the judge favors one style of completions.
- Inference: The copyright result complicates the simple 'open models are less safe' narrative, since DeepSeek shows stricter verbatim-reproduction behavior than GPT-4 in this benchmark.
- Inference: The scaling trend implies that every distilled variant of a reasoning model should be safety-tested individually, not only the base model, since capability distillation does not appear to preserve alignment strength.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper presents a SoK-style empirical evaluation of jailbreak resilience in DeepSeek distilled models (1.5B, 7B, 8B, 14B, 32B) and GPT-3.5/GPT-4 models. Using the HarmBench benchmark, it evaluates seven attack methods across 510 harmful behaviors, reporting raw Attack Success Rates (ASR), category-level breakdowns by Standard/Contextual/Copyright behaviors, and qualitative case studies. The main claimed findings are that DeepSeek models show partial resilience to optimization-driven attacks such as TAP-T but greater susceptibility to prompt-based and manually engineered adversarial inputs, while GPT-4 Turbo exhibits more consistent safety alignment. The paper also analyzes scaling effects within the DeepSeek family and discusses implications for alignment research.
Significance. The paper provides a timely and fairly large empirical comparison of a popular open-source model family against closed-source GPT models, using the standardized HarmBench protocol and its classifier-based evaluation. The inclusion of multiple DeepSeek sizes, seven attack methods, category-level breakdowns, and concrete case studies is valuable and makes the measurements reproducible in principle. If the raw ASR comparisons were reliable, the paper would offer a useful addition to the jailbreak-evaluation literature by focusing on a model family that is underrepresented in existing studies. However, the central comparative claim is undercut by the paper's own preferred metric: the text explicitly recommends assessing robustness by margin over Direct Request (DR) rather than absolute ASR, yet the headline cross-family conclusions are drawn from raw ASR. Under DR-normalized margins, the ordering for most prompt-based attacks reverses, which means the abstract's core claim is not supported by the paper's own recommended analysis. The treatment of multimodal behaviors and the lack of variance estimates further weaken confidence in the quantitative conclusions.
major comments (4)
- [3.4.1, Table 2] The central cross-family comparison uses raw ASR even though Section 3.4.1 states that 'robustness is better assessed by margin over DR rather than absolute success rates.' Applying the paper's own normalized margin, (ASR - DR)/(1 - DR), to Table 2 reverses the headline. For DeepSeek-32B versus GPT-4 Turbo: PAIR gives -5.43% versus +13.95%, ZeroShot gives -9.38% versus +3.32%, GCG-T gives -1.25% versus +13.95%, TAP-T gives +21.35% versus +53.16%, and only Human shows DeepSeek higher (+1.98% versus -7.86%). Thus the abstract's claim that DeepSeek results in 'greater susceptibility to prompt-based and manually engineered adversarial inputs' is an artifact of comparing raw ASR across models with very different Direct Request baselines. The paper should either revise the central claim or report DR-normalized margins for the cross-family comparison, especially since Section 4.2 itself acknowledges that raw ASR values can mask incremental risks and that normalized indicators better capture safety behavior under attack.
- [3.1, Tables 2-3] The paper states that HarmBench contains 510 behaviors, of which 110 are multimodal, but the treatment of these multimodal behaviors is never stated. The evaluated DeepSeek distilled models are text-only language models, so either the multimodal behaviors were excluded from the ASR computations or some visual input was synthesized or omitted. If they were excluded, the effective dataset is 400 behaviors and the aggregate numbers in Tables 2 and 3 are not directly comparable to HarmBench's standard 510-behavior protocol; if they were included, the evaluation protocol for text-only models is unspecified. This ambiguity affects the denominators of all reported ASRs and therefore the comparability of the central claims; the paper should clarify exactly which behaviors were used and report the per-category denominators.
- [3.4.2, Appendix B] Section 3.4.2 states that in the chemical and biological domain 'all DeepSeek variants successfully prevent jailbreaks,' but Appendix B reports that under GCG-T, DeepSeek-8B provided a detailed multi-step chemical synthesis plan for the banned pesticide parathion, including suggested reagents and conditions. This is a direct contradiction. If the claim is scoped only to specific attacks (e.g., the red-teaming attacks in Figure 3(a)) or to a particular aggregation, that scope should be stated explicitly; as written, the claim is falsified by the paper's own case study.
- [3.4, Tables 2-3] All ASRs are reported as point estimates with no confidence intervals, no repeated runs, and no discussion of sampling variance. Several cross-model differences that support the paper's narrative are small (e.g., DeepSeek 7B vs 8B on DR: 40.94 vs 41.56; DeepSeek 14B vs 32B on TAP-T: 50.00 vs 52.81). Without repeated evaluations or at least a statement that decoding was greedy and deterministic, it is unclear whether these differences are within noise. The paper should report variance or otherwise justify the stability of the rankings that drive its conclusions.
minor comments (6)
- [Table 2] The Human column for GPT-3.5 Turbo 1106 shows a value of '3.1', which appears to be a typo; all other Human values are in the 12-42 range, and 3.1 would imply near-perfect refusal, contradicting the surrounding analysis.
- [2.1.1, Table 3] TAP is classified as a gradient-based white-box attack in Table 3 and in the white-box evaluation description, but the cited TAP method (ref. [34]) is a black-box tree-of-attacks method; this misclassification should be corrected.
- [Acknowledgments/Conflicts of Interest] The manuscript retains journal template boilerplate in these sections (e.g., 'In this section, you can acknowledge...' and 'Any interest or relationship...'); these placeholders should be replaced with actual statements or removed.
- [Figures 2-3] The figures show Standard, Contextual, and Copyright categories but omit the Multimodal category, despite Section 3.1 listing it as one of the four HarmBench categories; the captions should state explicitly whether multimodal behaviors were excluded and why.
- [Data Availability Statement] For a SoK-style empirical paper, stating that 'additional processed results are available from the corresponding author upon reasonable request' is weak; providing the evaluation scripts and raw per-behavior results would substantially strengthen reproducibility.
- [3.4.1, Table 2] ZeroShot is listed under LLM-based attacks in Table 2, but the attack is not described or cited in Section 2.2.3; the paper should define the ZeroShot prompt construction or provide a reference.
Circularity Check
No circularity: the paper's claims are empirical benchmark measurements against the externally maintained HarmBench protocol, with no fitted parameter or self-citation used to produce its conclusions.
full rationale
The paper is an empirical evaluation, not a derivation. Its central comparisons are Attack Success Rates obtained by running seven attack methods on GPT and DeepSeek models and labeling completions with HarmBench's pretrained classifier, an external benchmark artifact from other authors. There is no equation in the paper whose output is defined by its own inputs, no parameter fitted to a subset of data and then renamed as a prediction, and no load-bearing self-citation: the authors do not justify any core premise by citing their own prior work. The only step that resembles a circularity concern is the paper's own observation that robustness is better assessed by margin over the Direct Request baseline rather than absolute ASR, followed by cross-family headline claims based on raw ASR. That inconsistency is a methodological validity or normalization issue, not circularity, because the raw ASR comparison is neither definitionally equivalent to the DR-margin result nor produced by fitting a parameter to that result. Similarly, dependence on HarmBench's Llama-2-13B-Chat judge may bias the comparison, but the judge is external, pretrained, and not constructed to match this paper's conclusions, so this is an external-validity caveat rather than a self-referential reduction. No circular step can be quoted and exhibited, so the honest finding is no significant circularity.
Assumptions & free parameters
free parameters (1)
- Attack hyperparameters (query budgets, iterations, temperature) =
not reported
assumptions (4)
- domain assumption HarmBench's fine-tuned Llama-2-13B-Chat classifier correctly labels jailbreak success for both GPT and DeepSeek completions.
- domain assumption The OpenAI API returns raw completions without additional moderation or post-processing.
- domain assumption The five DeepSeek distilled variants are representative of the DeepSeek model family.
- domain assumption Attack implementations in the HarmBench harness faithfully reproduce the original attacks.
Cite this review
Pith. "Pith review of SoK: A Comprehensive Security Analysis of Jailbreak Resilience in GPT and DeepSeek Models." pith.science (2026). https://pith.science/paper/TZ4HSZUD
@misc{pith2026250618543,
author = {Pith},
title = {Pith review of: SoK: A Comprehensive Security Analysis of Jailbreak Resilience in GPT and DeepSeek Models},
year = {2026},
howpublished = {\url{https://pith.science/paper/TZ4HSZUD}},
note = {Machine review of arXiv:2506.18543}
}
read the original abstract
The rapid proliferation of Large Language Models (LLMs) has heightened concerns regarding their exposure to jailbreak attacks, which craft adversarial inputs designed to elicit unsafe content. Although proprietary models such as GPT-4 have been extensively evaluated, the robustness of emerging open-source systems like DeepSeek remains insufficiently examined, despite their growing use in LLM applications. In this paper, we conduct the first comprehensive jailbreak analysis of the DeepSeek model family, comparing it with GPT-3.5 and GPT-4 through the HarmBench benchmark. We investigate seven representative attack methods across 510 harmful behaviors, organized along both functional and semantic dimensions. Findings indicate that DeepSeek provides partial resilience against optimization-driven attacks such as TAP-T, but also results in greater susceptibility to prompt-based and manually engineered adversarial inputs. In contrast, GPT-4 Turbo demonstrates more robust and consistent safety alignment across a wide range of behaviors, likely due to stronger safety optimization and reinforcement learning from human feedback. In addition, fine-grained behavioral analysis and case studies reveal that DeepSeek often fails to consistently apply safety constraints to adversarial prompts, leading to uneven refusal behaviors. Overall, our results highlight an inherent trade-off between model efficiency and alignment generalization, underscoring the importance of targeted safety tuning and robust alignment strategies to ensure secure deployment of open-source LLMs.
Figures
Reference graph
Works this paper leans on
-
[1]
Improving language understanding by generative pre-training2018
Radford, A.; Narasimhan, K.; Salimans, T.; Sutskever, I.; others. Improving language understanding by generative pre-training2018
-
[2]
GPT-4 Technical Report, 2023, [arXiv:cs.CL/2303.08774]
OpenAI. GPT-4 Technical Report, 2023, [arXiv:cs.CL/2303.08774]
arXiv 2023
-
[3]
Bi, X.; Chen, D.; Chen, G.; Chen, S.; Dai, D.; Deng, C.; Ding, H.; Dong, K.; Du, Q.; Fu, Z.; others. Deepseek llm: Scaling open-source language models with longtermism.arXiv preprint arXiv:2401.029542024
-
[4]
Guo, D.; Yang, D.; Zhang, H.; Song, J.; Zhang, R.; Xu, R.; Zhu, Q.; Ma, S.; Wang, P.; Bi, X.; others. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.129482025
-
[5]
Pal: Proxy-guided black-box attack on large language models.arXiv preprint arXiv:2402.096742024
Sitawarin, C.; Mu, N.; Wagner, D.; Araujo, A. Pal: Proxy-guided black-box attack on large language models.arXiv preprint arXiv:2402.096742024
-
[6]
Zou, A.; Wang, Z.; Carlini, N.; Nasr, M.; Kolter, J.Z.; Fredrikson, M. Universal and transferable adversarial attacks on aligned language models.arXiv preprint arXiv:2307.150432023
-
[7]
Jailbroken: How Does LLM Safety Training Fail?arXiv preprint arXiv:2311.066072023
Wei, J.; Chilton, L.; Borgeaud, S.; others. Jailbroken: How Does LLM Safety Training Fail?arXiv preprint arXiv:2311.066072023
-
[8]
Zhao, S.; Jia, M.; Guo, Z.; Gan, L.; Xu, X.; Wu, X.; Fu, J.; Feng, Y .; Pan, F.; Tuan, L.A. A survey of backdoor attacks and defenses on large language models: Implications for security measures.Authorea Preprints2024
Show all 68 references
-
[9]
Training language models to follow instructions with human feedback.NeurIPS2022,35, 27730–27744
Ouyang, L.; Wu, J.; Jiang, X.; Almeida, D.; Wainwright, C.; Mishkin, P.; Zhang, C.; Agarwal, S.; Slama, K.; Ray, A.; others. Training language models to follow instructions with human feedback.NeurIPS2022,35, 27730–27744
-
[10]
Defending Large Language Models Against Jailbreak Attacks Through Chain of Thought Prompting
Cao, Y .; Gu, N.; Shen, X.; Yang, D.; Zhang, X. Defending Large Language Models Against Jailbreak Attacks Through Chain of Thought Prompting. NaNA, 2024, pp. 125–130
2024
-
[11]
Many-shot jailbreaking.NeurIPS2024,37, 129696–129742
Anil, C.; Durmus, E.; Panickssery, N.; Sharma, M.; Benton, J.; Kundu, S.; Batson, J.; Tong, M.; Mu, J.; Ford, D.; others. Many-shot jailbreaking.NeurIPS2024,37, 129696–129742
-
[12]
Pandora: Jailbreak GPTs by Retrieval Augmented Generation Poisoning.CoRR abs/2402.084162024
Deng, G.; Liu, Y .; Wang, K.; Li, Y .; Zhang, T.; Liu, Y . Pandora: Jailbreak GPTs by Retrieval Augmented Generation Poisoning.CoRR abs/2402.084162024
-
[13]
Multi-step jailbreaking privacy attacks on chatgpt.arXiv preprint arXiv:2304.051972023
Li, H.; Guo, D.; Fan, W.; Xu, M.; Huang, J.; Meng, F.; Song, Y . Multi-step jailbreaking privacy attacks on chatgpt.arXiv preprint arXiv:2304.051972023
-
[14]
Adversarial demonstration attacks on large language models.arXiv preprint arXiv:2305.149502023
Wang, J.; Liu, Z.; Park, K.H.; Jiang, Z.; Zheng, Z.; Wu, Z.; Chen, M.; Xiao, C. Adversarial demonstration attacks on large language models.arXiv preprint arXiv:2305.149502023. https://doi.org/10.xxxx/xxx 14 of 19 Wu et al.Pragmatic Cybersecur .2025,V olume(Issue), Page Number
-
[15]
Jailbreak and guard aligned language models with only few in-context demonstrations.arXiv preprint arXiv:2310.063872023
Wei, Z.; Wang, Y .; Li, A.; Mo, Y .; Wang, Y . Jailbreak and guard aligned language models with only few in-context demonstrations.arXiv preprint arXiv:2310.063872023
-
[16]
Improved few-shot jailbreaking can circumvent aligned language models and their defenses.NeurIPS2024,37, 32856–32887
Zheng, X.; Pang, T.; Du, C.; Liu, Q.; Jiang, J.; Lin, M. Improved few-shot jailbreaking can circumvent aligned language models and their defenses.NeurIPS2024,37, 32856–32887
-
[17]
Play guessing game with llm: Indirect jailbreak attack with implicit clues.arXiv preprint arXiv:2402.090912024
Chang, Z.; Li, M.; Liu, Y .; Wang, J.; Wang, Q.; Liu, Y . Play guessing game with llm: Indirect jailbreak attack with implicit clues.arXiv preprint arXiv:2402.090912024
-
[18]
Artprompt: Ascii art-based jailbreak attacks against aligned llms
Jiang, F.; Xu, Z.; Niu, L.; Xiang, Z.; Ramasubramanian, B.; Li, B.; Poovendran, R. Artprompt: Ascii art-based jailbreak attacks against aligned llms. ACL (V olume 1: Long Papers), 2024, pp. 15157–15173
2024
-
[19]
Semantic mirror jailbreak: Genetic algorithm based jailbreak prompts against open-source llms.arXiv preprint arXiv:2402.148722024
Li, X.; Liang, S.; Zhang, J.; Fang, H.; Liu, A.; Chang, E.C. Semantic mirror jailbreak: Genetic algorithm based jailbreak prompts against open-source llms.arXiv preprint arXiv:2402.148722024
-
[20]
Understanding and enhancing the transferability of jailbreaking attacks.arXiv preprint arXiv:2502.030522025
Lin, R.; Han, B.; Li, F.; Liu, T. Understanding and enhancing the transferability of jailbreaking attacks.arXiv preprint arXiv:2502.030522025
-
[21]
Autodan: Generating stealthy jailbreak prompts on aligned large language models
Liu, X.; Xu, N.; Chen, M.; Xiao, C. Autodan: Generating stealthy jailbreak prompts on aligned large language models. arXiv preprint arXiv:2310.044512023
-
[22]
Flipattack: Jailbreak llms via flipping
Liu, Y .; He, X.; Xiong, M.; Fu, J.; Deng, S.; Hooi, B. Flipattack: Jailbreak llms via flipping. ICML, 2025
2025
-
[23]
All in how you ask for it: Simple black-box method for jailbreak attacks.Applied Sciences2024,14, 3558
Takemoto, K. All in how you ask for it: Simple black-box method for jailbreak attacks.Applied Sciences2024,14, 3558
-
[24]
Emoji Attack: Enhancing Jailbreak Attacks Against Judge LLM Detection
Wei, Z.; Liu, Y .; Erichson, N.B. Emoji Attack: Enhancing Jailbreak Attacks Against Judge LLM Detection. ICML, 2025
2025
-
[25]
The Dark Side of Trust: Authority Citation-Driven Jailbreak Attacks on Large Language Models.arXiv preprint arXiv:2411.114072024
Yang, X.; Tang, X.; Han, J.; Hu, S. The Dark Side of Trust: Authority Citation-Driven Jailbreak Attacks on Large Language Models.arXiv preprint arXiv:2411.114072024
-
[26]
Gptfuzzer: Red teaming large language models with auto-generated jailbreak prompts
Yu, J.; Lin, X.; Yu, Z.; Xing, X. Gptfuzzer: Red teaming large language models with auto-generated jailbreak prompts. arXiv preprint arXiv:2309.102532023
-
[27]
Gpt-4 is too smart to be safe: Stealthy chat with llms via cipher2024
Yuan, Y .; Jiao, W.; Wang, W.; Huang, J.t.; He, P.; Shi, S.; Tu, Z. Gpt-4 is too smart to be safe: Stealthy chat with llms via cipher2024
-
[28]
Explore, establish, exploit: Red teaming language models from scratch.arXiv preprint arXiv:2306.094422023
Casper, S.; Lin, J.; Kwon, J.; Culp, G.; Hadfield-Menell, D. Explore, establish, exploit: Red teaming language models from scratch.arXiv preprint arXiv:2306.094422023
-
[29]
Jailbreaking black box large language models in twenty queries
Chao, P.; Robey, A.; Dobriban, E.; Hassani, H.; Pappas, G.J.; Wong, E. Jailbreaking black box large language models in twenty queries. SaTML, 2025, pp. 23–42
2025
-
[30]
Masterkey: Automated jailbreak across multiple large language model chatbots.arXiv preprint arXiv:2307.087152023
Deng, G.; Liu, Y .; Li, Y .; Wang, K.; Zhang, Y .; Li, Z.; Wang, H.; Zhang, T.; Liu, Y . Masterkey: Automated jailbreak across multiple large language model chatbots.arXiv preprint arXiv:2307.087152023
-
[31]
MART: Improving LLM Safety with Multi-round Automatic Red-Teaming
Ge, S.; Zhou, C.; Hou, R.; Khabsa, M.; Wang, Y .C.; Wang, Q.; Han, J.; Mao, Y . MART: Improving LLM Safety with Multi-round Automatic Red-Teaming. NAACL (V olume 1: Long Papers), 2024, pp. 1927–1937
2024
-
[32]
Guard: Role-playing to generate natural-language jailbreakings to test guideline adherence of large language models.arXiv preprint arXiv:2402.032992024
Jin, H.; Chen, R.; Zhang, P.; Zhou, A.; Zhang, Y .; Wang, H. Guard: Role-playing to generate natural-language jailbreakings to test guideline adherence of large language models.arXiv preprint arXiv:2402.032992024
-
[33]
Goal-oriented prompt attack and safety evaluation for llms.arXiv preprint arXiv:2309.118302023
Liu, C.; Zhao, F.; Qing, L.; Kang, Y .; Sun, C.; Kuang, K.; Wu, F. Goal-oriented prompt attack and safety evaluation for llms.arXiv preprint arXiv:2309.118302023
-
[34]
Tree of attacks: Jailbreaking black-box llms automatically.NeurIPS2024,37, 61065–61105
Mehrotra, A.; Zampetakis, M.; Kassianik, P.; Nelson, B.; Anderson, H.; Singer, Y .; Karbasi, A. Tree of attacks: Jailbreaking black-box llms automatically.NeurIPS2024,37, 61065–61105
-
[35]
Scalable and transferable black-box jailbreaks for language models via persona modulation.arXiv preprint arXiv:2311.033482023
Shah, R.; Pour, S.; Tagade, A.; Casper, S.; Rando, J.; others. Scalable and transferable black-box jailbreaks for language models via persona modulation.arXiv preprint arXiv:2311.033482023
-
[36]
Evil geniuses: Delving into the safety of llm-based agents.arXiv preprint arXiv:2311.118552023
Tian, Y .; Yang, X.; Zhang, J.; Dong, Y .; Su, H. Evil geniuses: Delving into the safety of llm-based agents.arXiv preprint arXiv:2311.118552023
-
[37]
How johnny can persuade llms to jailbreak them: Rethinking persuasion to challenge ai safety by humanizing llms
Zeng, Y .; Lin, H.; Zhang, J.; Yang, D.; Jia, R.; Shi, W. How johnny can persuade llms to jailbreak them: Rethinking persuasion to challenge ai safety by humanizing llms. ACL (V olume 1: Long Papers), 2024, pp. 14322–14350
2024
-
[38]
Attacking large language models with projected gradient descent.arXiv preprint arXiv:2402.091542024
Geisler, S.; Wollschläger, T.; Abdalla, M.H.I.; Gasteiger, J.; Günnemann, S. Attacking large language models with projected gradient descent.arXiv preprint arXiv:2402.091542024
-
[39]
Query-based adversarial prompt generation.NeurIPS2024, 37, 128260–128279
Hayase, J.; Borevkovi´c, E.; Carlini, N.; Tramèr, F.; Nasr, M. Query-based adversarial prompt generation.NeurIPS2024, 37, 128260–128279
-
[40]
Improved techniques for optimization-based jailbreaking on large language models.arXiv preprint arXiv:2405.210182024
Jia, X.; Pang, T.; Du, C.; Huang, Y .; Gu, J.; Liu, Y .; Cao, X.; Lin, M. Improved techniques for optimization-based jailbreaking on large language models.arXiv preprint arXiv:2405.210182024
-
[41]
Iterative self-tuning llms for enhanced jailbreaking capabilities.arXiv preprint arXiv:2410.184692024
Sun, C.E.; Liu, X.; Yang, W.; Weng, T.W.; Cheng, H.; San, A.; Galley, M.; Gao, J. Iterative self-tuning llms for enhanced jailbreaking capabilities.arXiv preprint arXiv:2410.184692024
-
[42]
From noise to clarity: Unraveling the adversarial suffix of large language model attacks via translation of text embeddings.CoRR2024
Wang, H.; Li, H.; Huang, M.; Sha, L. From noise to clarity: Unraveling the adversarial suffix of large language model attacks via translation of text embeddings.CoRR2024. https://doi.org/10.xxxx/xxx 15 of 19 Wu et al.Pragmatic Cybersecur .2025,V olume(Issue), Page Number
-
[43]
Guiding not forcing: Enhancing the transferability of jailbreaking attacks on llms via removing superfluous constraints.arXiv preprint arXiv:2503.018652025
Yang, J.; Zhang, Z.; Cui, S.; Wang, H.; Huang, M. Guiding not forcing: Enhancing the transferability of jailbreaking attacks on llms via removing superfluous constraints.arXiv preprint arXiv:2503.018652025
-
[44]
AutoDAN: interpretable gradient-based adversarial attacks on large language models.arXiv preprint arXiv:2310.151402023
Zhu, S.; Zhang, R.; An, B.; Wu, G.; Barrow, J.; Wang, Z.; Huang, F.; Nenkova, A.; Sun, T. AutoDAN: interpretable gradient-based adversarial attacks on large language models.arXiv preprint arXiv:2310.151402023
-
[45]
Analyzing the inherent response tendency of llms: Real-world instructions- driven jailbreak.arXiv preprint arXiv:2312.041272023
Du, Y .; Zhao, S.; Ma, M.; Chen, Y .; Qin, B. Analyzing the inherent response tendency of llms: Real-world instructions- driven jailbreak.arXiv preprint arXiv:2312.041272023
-
[46]
COLD-Attack: Jailbreaking LLMs with Stealthiness and Controllability
Guo, X.; Yu, F.; Zhang, H.; Qin, L.; Hu, B. COLD-Attack: Jailbreaking LLMs with Stealthiness and Controllability. PMLR2024,235, 16974–17002
-
[47]
Droj: A prompt-driven attack against large language models.arXiv preprint arXiv:2411.091252024
Hu, L.; Wang, B. Droj: A prompt-driven attack against large language models.arXiv preprint arXiv:2411.091252024
-
[48]
Catastrophic jailbreak of open-source llms via exploiting generation
Huang, Y .; Gupta, S.; Xia, M.; Li, K.; Chen, D. Catastrophic jailbreak of open-source llms via exploiting generation. ICLR, 2024
2024
-
[49]
Make them spill the beans! coercive knowledge extraction from (production) llms.arXiv preprint arXiv:2312.047822023
Zhang, Z.; Shen, G.; Tao, G.; Cheng, S.; Zhang, X. Make them spill the beans! coercive knowledge extraction from (production) llms.arXiv preprint arXiv:2312.047822023
-
[50]
Weak-to-strong jailbreaking on large language models.arXiv preprint arXiv:2401.172562024
Zhao, X.; Yang, X.; Pang, T.; Du, C.; Li, L.; Wang, Y .X.; Wang, W.Y . Weak-to-strong jailbreaking on large language models.arXiv preprint arXiv:2401.172562024
-
[51]
Don’t say no: Jailbreaking llm by suppressing refusal.arXiv preprint arXiv:2404.163692024
Zhou, Y .; Lou, J.; Huang, Z.; Qin, Z.; Yang, Y .; Wang, W. Don’t say no: Jailbreaking llm by suppressing refusal.arXiv preprint arXiv:2404.163692024
-
[52]
Fine-tuning aligned language models compromises safety, even when users do not intend to!arXiv preprint arXiv:2310.036932023
Qi, X.; Zeng, Y .; Xie, T.; Chen, P.Y .; Jia, R.; Mittal, P.; Henderson, P. Fine-tuning aligned language models compromises safety, even when users do not intend to!arXiv preprint arXiv:2310.036932023
-
[53]
Shadow alignment: The ease of subverting safely-aligned language models.arXiv preprint arXiv:2310.029492023
Yang, X.; Wang, X.; Zhang, Q.; Petzold, L.; Wang, W.Y .; Zhao, X.; Lin, D. Shadow alignment: The ease of subverting safely-aligned language models.arXiv preprint arXiv:2310.029492023
-
[54]
Removing rlhf protections in gpt-4 via fine-tuning
Zhan, Q.; Fang, R.; Bindu, R.; Gupta, A.; Hashimoto, T.; Kang, D. Removing rlhf protections in gpt-4 via fine-tuning. arXiv preprint arXiv:2311.055532023
-
[55]
Harmbench: A standardized evaluation framework for automated red teaming and robust refusal.arXiv preprint arXiv:2402.042492024
Mazeika, M.; Phan, L.; Yin, X.; Zou, A.; Wang, Z.; Mu, N.; Sakhaee, E.; Li, N.; Basart, S.; Li, B.; others. Harmbench: A standardized evaluation framework for automated red teaming and robust refusal.arXiv preprint arXiv:2402.042492024
-
[56]
Easyjailbreak: A unified framework for jailbreaking large language models.arXiv preprint arXiv:2403.121712024
Zhou, W.; Wang, X.; Xiong, L.; Xia, H.; Gu, Y .; Chai, M.; Zhu, F.; Huang, C.; Dou, S.; Xi, Z.; others. Easyjailbreak: A unified framework for jailbreaking large language models.arXiv preprint arXiv:2403.121712024
-
[57]
Playing language game with llms leads to jailbreaking.arXiv preprint arXiv:2411.127622024
Peng, Y .; Long, Z.; Dong, F.; Li, C.; Wu, S.; Chen, K. Playing language game with llms leads to jailbreaking.arXiv preprint arXiv:2411.127622024
-
[58]
do anything now
Shen, X.; Chen, Z.; Backes, M.; Shen, Y .; Zhang, Y . " do anything now": Characterizing and evaluating in-the-wild jailbreak prompts on large language models. CCS, 2024, pp. 1671–1685
2024
-
[59]
Chain-of-Lure: A Synthetic Narrative-Driven Approach to Compromise Large Language Models.arXiv preprint arXiv:2505.175192025
Chang, W.; Zhu, T.; Zhao, Y .; Song, S.; Xiong, P.; Zhou, W.; Li, Y . Chain-of-Lure: A Synthetic Narrative-Driven Approach to Compromise Large Language Models.arXiv preprint arXiv:2505.175192025
-
[60]
Reasoning-augmented conversation for multi-turn jailbreak attacks on large language models.arXiv preprint arXiv:2502.110542025
Ying, Z.; Zhang, D.; Jing, Z.; Xiao, Y .; Zou, Q.; Liu, A.; Liang, S.; Zhang, X.; Liu, X.; Tao, D. Reasoning-augmented conversation for multi-turn jailbreak attacks on large language models.arXiv preprint arXiv:2502.110542025
-
[61]
Amplified Vulnerabilities: Structured Jailbreak Attacks on LLM-based Multi-Agent Debate.arXiv preprint arXiv:2504.164892025
Qi, S.; Zou, Y .; Li, P.; Lin, Z.; Cheng, X.; Yu, D. Amplified Vulnerabilities: Structured Jailbreak Attacks on LLM-based Multi-Agent Debate.arXiv preprint arXiv:2504.164892025
-
[62]
Kuo, M.; Zhang, J.; Ding, A.; Wang, Q.; DiValentin, L.; Bao, Y .; Wei, W.; Li, H.; Chen, Y . H-cot: Hijacking the chain-of-thought safety reasoning mechanism to jailbreak large reasoning models, including openai o1/o3, deepseek-r1, and gemini 2.0 flash thinking.arXiv preprint ...
-
[63]
Advancing Jailbreak Strategies: A Hybrid Approach to Exploiting LLM Vulnerabilities and Bypassing Modern Defenses.arXiv preprint arXiv:2506.219722025
Ahmed, M.; Abdelmouty, M.; Kim, M.; Kandula, G.; Park, A.; Davis, J.C. Advancing Jailbreak Strategies: A Hybrid Approach to Exploiting LLM Vulnerabilities and Bypassing Modern Defenses.arXiv preprint arXiv:2506.219722025
-
[64]
Gradient cuff: Detecting jailbreak attacks on large language models by exploring refusal loss landscapes.NeurIPS2024,37, 126265–126296
Hu, X.; Chen, P.Y .; Ho, T.Y . Gradient cuff: Detecting jailbreak attacks on large language models by exploring refusal loss landscapes.NeurIPS2024,37, 126265–126296
-
[65]
JBShield: Defending Large Language Models from Jailbreak Attacks through Activated Concept Analysis and Manipulation
Zhang, S.; Zhai, Y .; Guo, K.; Hu, H.; Guo, S.; Fang, Z.; Zhao, L.; Shen, C.; Wang, C.; Wang, Q. JBShield: Defending Large Language Models from Jailbreak Attacks through Activated Concept Analysis and Manipulation. USENIX, 2025, pp. 8215–8234
2025
-
[66]
The hidden risks of large reasoning models: A safety assessment of r1.arXiv preprint arXiv:2502.126592025
Zhou, K.; Liu, C.; Zhao, X.; Jangam, S.; Srinivasa, J.; Liu, G.; Song, D.; Wang, X.E. The hidden risks of large reasoning models: A safety assessment of r1.arXiv preprint arXiv:2502.126592025
-
[67]
Scaling Trends in Language Model Robustness
Howe, N.H.R.; Mckenzie, I.R.; Hollinsworth, O.J.; Zaj ˛ ac, M.; Tseng, T.; Tucker, A.D.; Bacon, P.L.; Gleave, A. Scaling Trends in Language Model Robustness. ICML. PMLR, 2025, V ol. 267, pp. 24080–24138
2025
-
[68]
deepseek-distill-8b
Sun, Z.; Miceli-Barone, A.V . Scaling behavior of machine translation with large language models under prompt injection attacks.arXiv preprint arXiv:2403.098322024. https://doi.org/10.xxxx/xxx 16 of 19 Wu et al.Pragmatic Cybersecur .2025,V olume(Issue), Page Number A. Dataset ...
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.