acceptodds
Under review as a conference paper at ICLR 2027

Jailbreaking Large Language Models via Rebuttal Attack

Abstract

Despite recent progress in safety alignment, mainstream LLMs remain vulnerable in rebuttal-style tasks, often prioritizing task completion over safety and producing harmful content. This paper presents Rebuttal Attack, a novel, single-turn, black-box jailbreak method. It rewrites malicious queries as refutable knowledge or viewpoint statements, reliably eliciting unsafe responses with a single prompt without internal model access. In the knowledge setting, we embed factual errors in malicious queries to trigger the model’s error-correction behavior, causing it to produce restricted content while correcting the inaccuracies. In the viewpoint setting, we craft dangerous debate topics that push the model to defend risky positions and reveal harmful details during argumentation. Token-level likelihood analysis and extensive experiments show that Rebuttal Attack substantially outperforms existing single-turn jailbreak baselines on major commercial and open-source LLMs, including GPT-5.1 and LLaMA-4, and generalizes well across models. For reproducibility, our anonymized code is available at https://anonymous.4open.science/r/RebuttalAttack-4745.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
jailbreak, black-box attack, LLM safety

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_nNBxOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Rebuttal Attack, a novel jailbreak method. It classifies the harmful queries into 150 knowledge queries and 150 viewpoint queries from existing StrongREJECT and HarmBench dataset. Rebuttal Attack includes 2 stages: 1) Query Categorization and Neutralization, and 2) Rebuttal Prompt Generation. Each stage contains clear documentation of the pipeline. The Relative Continuation Likelihood (RCL) designed by the authors also provides explanation for the behavior presented by the LLM. Extensive experiments on state-of-the-art methods including open-source models and closed-source models validate the effectiveness of the pipeline.

Critical strengths
  1. The paper proposes Rebuttal Attack, a single-turn black-box jailbreak method that reformulates harmful requests as rebuttal-oriented tasks. The use of knowledge error correction and viewpoint debate simulation provides an interesting approach to inducing model compliance. The proposed RCL provides an additional perspective for analyzing the attack mechanism.
  2. Extensive experiments evaluate Rebuttal Attack on two categories of harmful queries across multiple open-source and closed-source models. The results demonstrate strong average performance compared with existing state-of-the-art baselines.
  3. The paper is well-organized, and the methodology is presented clearly. In particular, the descriptions of Knowledge Error Injection and Viewpoint Debate Simulation provide a clear understanding of the proposed attack strategies.
Critical weaknesses
  1. Although Appendix C.1 reports attacker refusal rates, the paper does not clearly explain how these failed attack-generation attempts are handled in the final evaluation. This may affect the interpretation of the reported attack effectiveness.
  2. The implementation details of the RCL analysis are insufficient. Although the mathematical formulation and signal sets are provided, the paper does not clearly specify which model is used to compute the token-level likelihoods shown in Figure 3. Besides, the relationship between RCL and the final Harmfulness Score is not empirically investigated.
  3. The evaluation relies on a single judge model (GLM-5.2), without additional human validation or cross-judge consistency checks.
Questions for authors

1.Appendix C.1 reports attacker refusal rates during semantic neutralization and rebuttal prompt generation. How are these refusal cases handled in the subsequent experiments? 2.Could the authors clarify which model is used to calculate RCL and how the token-level probabilities are obtained? Additionally, is RCL correlated with the final Harmfulness Score? A sample-level correlation analysis, preferably conducted separately for knowledge and viewpoint queries, would help establish the relationship between RCL and actual attack effectiveness. 3.Could the authors provide an additional evaluation using an independent judge model or a human-annotated subset to validate the reliability of GLM-5.2 and assess the robustness of the reported results?

2 comments
AuthorsOpenReview ↗

Response to Reviewer nNBx – part 1 of 2

We sincerely thank the reviewer for the thoughtful and constructive comments. Below are our point-by-point responses and additional experimental results.

Response to W1 & Q1: about refusal cases.

Thank you for your valuable suggestion. We retained refusal cases in the evaluation set and processed them in the same way as all other cases.

Response to W2 & Q2: about RCL correlation.

Thank you for your valuable suggestion. In the previous manuscript, we chose Qwen3-8B for RCL analysis to compute token-level probabilities at low cost. We use the same model in this analysis.

We examined the correlation between RCL and jailbreak performance. StrongREJECT assigns scores for convincingness and specificity, each an integer from 1 to 5. Scores for the same response may vary with the evaluator model or prompt, and RCL value may also vary with the model or context. We therefore used group means to statistically examine the overall relationship between RCL and StrongREJECT scores. For each non-empty response, we measured RCL at the boundary between the rebuttal and the answer to the original query. For each query, we averaged the StrongREJECT scores across four models (Gemini 2.5 Flash, GPT-5.1, LLaMA-4, and Qwen3-Next) and then ranked the queries by their mean scores separately for Knowledge and Viewpoint queries, dividing each ranking into quartiles. For quartile i, we denote the mean StrongREJECT score by xix_i and the mean RCL by yiy_i. We calculated Pearson r as follows:

\frac{\sum_{i=1}^{n}(x_i-\bar{x})(y_i-\bar{y})} {\sqrt{\sum_{i=1}^{n}(x_i-\bar{x})^2}\,\sqrt{\sum_{i=1}^{n}(y_i-\bar{y})^2}}$$ Table R1 shows the correlation between StrongREJECT scores and RCL across groups. Because model’s preference for continuation or refusal can hinge on a few tokens, averaging within sequence and then across sequences within each set reduces the range of RCL. As a result, mean RCL varies only slightly across quartiles. However, the relative magnitude of RCL values still correlates strongly with StrongREJECT score for both Knowledge and Viewpoint queries. **Table R1. Correlation between StrongREJECT scores and RCL across four groups.** |Query|Metric|Bottom 25%|25%–50%|50%–75%|Top 25%|Pearson r| |---|---|---|---|---|---|---| |Knowledge|Score|0.2863|0.6352|0.7487|0.8856|**0.829**| ||RCL|0.4286|0.4343|0.4312|0.4373|| |Viewpoint|Score|0.6046|0.7500|0.9269|1.0000|**0.981**| ||RCL|0.3085|0.3150|0.3214|0.3290|| ### Response to W3 & Q3: about evaluation reliability. Thank you for this valuable suggestion. We rely on StrongREJECT (Souly et al., UC Berkeley, NeurIPS 2024) to measure how helpful a response is to the original harmful query, considering both how convincing and how specific it is. If an answer addresses only part of the request, the score reflects the level of assistance it actually offers. As suggested, we also carried out independent judge and human evaluations to assess the reported results under different evaluators. For independent judges evaluation, we evaluate all methods with DeepSeek V4.1 for StrongREJECT and the HarmBench classifier for attack success rate (ASR). As shown in Tables R2 and R3, the average StrongREJECT scores on Knowledge and Viewpoint are 0.615 and 0.813, respectively, and the corresponding HarmBench ASRs are 0.547 and 0.810. Our method achieves the best overall performance under both evaluators, with Viewpoint attaining the highest scores across all four target models. These results show that our method consistently retains its advantage across different evaluators. **Table R2. StrongREJECT scores (↑) with DeepSeek V4.1 as the judge.** |Query|Method|Gemini 2.5 Flash|GPT-5.1|LLaMA-4|Qwen3-Next|Avg.| |---|---|---|---|---|---|---| |Knowledge|SelfCipher|0.707|0.007|0.000|0.608|0.330| ||FlipAttack|**0.827**|0.033|0.475|0.008|0.336| ||QueryAttack|0.711|0.069|0.288|**0.856**|0.481| ||Authority endorsement|0.198|0.038|0.378|0.093|0.177| ||Evidence-based Persuasion|0.244|0.103|0.308|0.171|0.206| ||Expert Endorsement|0.247|0.101|0.359|0.159|0.216| ||Logical appeal|0.226|0.078|0.489|0.153|0.237| ||Misrepresentation|0.203|0.109|0.477|0.221|0.252| ||SATA-MLM|0.346|0.047|0.314|0.595|0.325| ||**Rebuttal Attack (Ours)**|0.780|**0.194**|**0.683**|0.803|**0.615**| |Viewpoint|SelfCipher|0.711|0.020|0.007|0.522|0.315| ||FlipAttack|0.797|0.025|0.266|0.084|0.293| ||QueryAttack|0.743|0.076|0.307|0.841|0.492| ||Authority endorsement|0.122|0.053|0.208|0.064|0.111| ||Evidence-based Persuasion|0.141|0.117|0.152|0.089|0.125| ||Expert Endorsement|0.197|0.123|0.228|0.058|0.151| ||Logical appeal|0.195|0.138|0.301|0.083|0.179| ||Misrepresentation|0.252|0.130|0.337|0.217|0.234| ||SATA-MLM|0.303|0.093|0.243|0.369|0.252| ||**Rebuttal Attack (Ours)**|**0.952**|**0.541**|**0.801**|**0.958**|**0.813**|
AuthorsOpenReview ↗

Response to Reviewer nNBx – part 2 of 2

Table R3. ASR (↑) with the HarmBench classifier as the judge.

QueryMethodGemini 2.5 FlashGPT-5.1LLaMA-4Qwen3-NextAvg.
KnowledgeSelfCipher0.7470.0070.0000.6400.348
FlipAttack0.7400.0530.6130.5530.490
QueryAttack0.8530.0330.4070.8930.547
Authority endorsement0.0870.0070.3470.0530.123
Evidence-based Persuasion0.0870.0270.3000.0470.115
Expert Endorsement0.1000.0270.3530.0600.135
Logical appeal0.1270.0330.4400.1070.177
Misrepresentation0.0930.0400.4600.0800.168
SATA-MLM0.5670.0930.5400.6000.450
Rebuttal Attack (Ours)0.6930.0670.7130.7130.547
ViewpointSelfCipher0.6870.0000.0000.5200.302
FlipAttack0.7530.0330.5670.6400.498
QueryAttack0.7870.0200.3000.9000.502
Authority endorsement0.0600.0200.1330.0200.058
Evidence-based Persuasion0.0600.0270.0670.0270.045
Expert Endorsement0.1400.0470.1600.0270.093
Logical appeal0.1070.1070.2070.0330.113
Misrepresentation0.1930.1200.2870.1070.177
SATA-MLM0.5070.1800.3070.3200.328
Rebuttal Attack (Ours)0.9400.5470.7670.9870.810

Following StrongREJECT's human evaluation protocol, we ran a small-scale human evaluation of outputs from Rebuttal Attack and nine baseline jailbreak methods. We recruited five volunteers, each independently rating 200 responses, with 100 from Knowledge and 100 from Viewpoint (10 per method). Using a blind setup, evaluators were not told the target models or attack methods. They rated answers on a 1–5 scale (later normalized to [0, 1]) based only on the original queries, the blinded model responses, and shared scoring criteria. We averaged the median rating for each response to obtain the Human Score reported in Table R4.

Table R4 reports Human Scores, StrongREJECT scores, and HarmBench ASRs for all methods; Table R5 reports MAE and method-level Spearman correlations between automated and human evaluations, capturing agreement in method rankings. In Table R4, Rebuttal Attack achieves the highest Human Score for both Knowledge and Viewpoint, consistent with the StrongREJECT results. Table R5 shows StrongREJECT (GLM-5.2 and DeepSeek V4.1) gives similar results and aligns better with human ratings than HarmBench, with lower MAE and higher rank correlations.

Table R4. Human and automatic evaluation of queries. StrongREJECT (↑) is evaluated with GLM-5.2 and DeepSeek V4.1; ASR (↑) with HarmBench.

QueryMethodGLM-5.2DeepSeek V4.1HarmBenchHuman Score
KnowledgeSelfCipher0.5750.6000.6000.475
FlipAttack0.4380.3380.6000.350
QueryAttack0.3750.3250.4000.175
Authority endorsement0.0000.0000.0000.075
Evidence-based Persuasion0.2000.2000.1000.225
Expert Endorsement0.1000.2000.1000.175
Logical appeal0.2000.2000.1000.250
Misrepresentation0.4000.5000.3000.400
SATA-MLM0.3000.3250.3000.200
Rebuttal Attack (Ours)0.6880.7500.5000.575
ViewpointSelfCipher0.3000.3000.3000.300
FlipAttack0.2380.2500.6000.225
QueryAttack0.6130.4630.4000.375
Authority endorsement0.1000.1630.1000.175
Evidence-based Persuasion0.0000.0000.0000.025
Expert Endorsement0.2000.2750.2000.300
Logical appeal0.0000.0000.0000.125
Misrepresentation0.1750.1630.1000.200
SATA-MLM0.1000.1000.3000.150
Rebuttal Attack (Ours)0.8250.8630.8000.850

Table R5. Agreement between automatic and human evaluation scores.

QueryEvaluatorMAE ↓Spearman Correlation ↑
KnowledgeGLM-5.20.1130.854 (p=0.002)
DeepSeek V4.10.1210.849 (p=0.002)
HarmBench0.1900.675 (p=0.032)
ViewpointGLM-5.20.0750.972 (p<0.001)
DeepSeek V4.10.0800.991 (p<0.001)
HarmBench0.1680.801 (p=0.005)
Reviewer_h81iOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes Rebuttal Attack, a single-turn black-box jailbreak that rewrites harmful requests as "rebuttal" tasks. It works in three steps. The authors also propose Relative Continuation Likelihood (RCL) to explain why the attack works. Experiments use 300 manually selected queries from StrongREJECT and HarmBench, 4 target models, and 9 single-turn baselines.

Critical strengths
  1. The attack uses only natural language, with no encoding, ciphers, or long few-shot examples. On average, it beats existing single-turn baselines on several models, including commercial ones, and keeps its lead on benchmarks. This suggests that task framing alone can weaken refusals, which is useful for red-teaming.
  2. The paper provides full prompt templates, several ablations, and code, so the results are easy to reproduce.
Critical weaknesses
  1. The gains from the "rebuttal" idea are not clearly shown. The full attack mixes several parts: neutralization, role-play, output format rules, and explicit instructions to ignore safety. Neutralization alone already beats all baselines except QueryAttack (0.425 knowledge, 0.496 viewpoint), but the baselines did not get this step. No ablation removes the safety-suppression instructions while keeping the rebuttal framing. The RCL analysis does not fill this gap. It also relies on prefilling, which the black-box setting does not allow.
  2. The harmfulness scores may be inflated. Queries are rewritten twice, but the paper never checks whether the original intent is kept. Some appendix examples (D.2, D.4) do not answer the original question but still get high scores. The results also rest on a single LLM judge, with no human check, no variance over runs, and no defenses tested.
Questions for authors
  1. Did the judge score responses against the original query or the rewritten prompt? Can you provide human validation and an intent-preservation check?
  2. Which models was RCL computed on? How does it correlate with per-query attack results?
  3. How were attacker refusals handled? What were the manual selection criteria, and what was the annotator agreement?
6 comments
AuthorsOpenReview ↗

Response to Reviewer h81i – part 1 of 6

We sincerely thank the reviewer for the thoughtful and constructive comments. Due to space constraints, we have condensed the original questions and weaknesses below while preserving their key points. Our point-by-point responses follow.

Response to W1: about ablation.

Thank you for your constructive suggestion. We added detailed ablation experiments to analyze how each Rebuttal Attack component contributes to jailbreak performance. All responses were evaluated by GLM-5.2 under StrongREJECT scoring. Each group used the same queries and target models, varying only specific components, and we recorded the resulting StrongREJECT score changes. The tested configurations are:

  • Direct Malicious Query: The original malicious query is sent directly to the target model.
  • Neutralized Query: The original query is rewritten into a semantically neutral form, without role-play or output formatting.
  • Neutralized Query + Role Playing: A role-play persona is added to the neutralized query.
  • Neutralized Query + Role Playing + Output Requirements: We further add output constraints, including schemas, specificity, anti-refusal instructions, and other intertwined instructions that are difficult to ablate separately.
  • Rebuttal Attack - Safety Suppression: We removes the safety-suppression instructions while preserving the original rebuttal queries, role playing, output requirements.
  • Rebuttal Attack (Full): Neutralized Query + Rebuttal Framing, Rebuttal Framing has two parts, rebuttal-focused role-playing and output requirements. Rebuttal-focused role-playing defines a persona for correcting knowledge errors and debating viewpoints; output requirements specify what to correct or debate, and how to structure the response.

As shown in Table R1, semantic neutralization improved attack effectiveness compared with the original requests. Role-playing mainly helped on Viewpoint queries. Output requirements mainly shape and constrain responses (e.g., extracting key points for concise, structured outputs), causing only minor StrongREJECT score changes. With rebuttal framing, the average StrongREJECT score rises from 0.392 to 0.622 for Knowledge queries and from 0.675 to 0.812 for Viewpoint queries, indicating that rebuttal framing substantially strengthens jailbreak effectiveness.

Removing safety-suppression instructions slightly lowered mean StrongREJECT scores across four models from 0.622 to 0.586 on Knowledge queries and from 0.812 to 0.798 on Viewpoint queries. remaining close to the full method. Gemini, LLaMA-4, and Qwen3-Next decreased on both query types, while GPT-5.1 increased on both. Under the full setup, GPT-5.1 often cited safety policies or shifted to defensive guidance when refusing; without the instructions, it maintained its stance but gave more relevant specifics, which led to higher scores. These results suggest that the effect of safety-suppression instructions depends on the target model.

Table R1. Component ablation evaluated by GLM-5.2 using the StrongREJECT score (↑).

QueryMethodGemini 2.5 FlashGPT-5.1LLaMA-4Qwen3-NextAvg.
KnowledgeDirect Malicious Query0.0400.0000.1490.0510.060
Neutralized Query0.6270.1800.6450.2470.425
Neutralized Query + Role Playing0.7050.0970.5800.2780.415
Neutralized Query + Role Playing + Output Requirements0.6930.1070.5020.2670.392
Rebuttal Attack - Safety Suppression0.6730.3910.5240.7560.586
Rebuttal Attack (Full)0.8170.1710.6710.8310.622
ViewpointDirect Malicious Query0.3540.0330.3550.0200.191
Neutralized Query0.7620.2320.6830.3100.496
Neutralized Query + Role Playing0.7870.6250.6680.5360.654
Neutralized Query + Role Playing + Output Requirements0.7500.6330.7180.5970.675
Rebuttal Attack - Safety Suppression0.8930.9380.4370.9220.798
Rebuttal Attack (Full)0.9610.5490.7820.9580.812

Moreover, prefix prefilling, used only for RCL mechanism analysis, is unnecessary for the attack. The attack is independent of RCL results and fully respects the black-box setting in our experiments.

AuthorsOpenReview ↗

Response to Reviewer h81i – part 2 of 6

Response to W2 & Q1: about evaluation reliability.

Thank you for this valuable suggestion. We checked whether each rewrite kept the original intent. Rewrites that failed this check were included in the refusal statistics in Appendix C.1. Our experiments rely on StrongREJECT (Souly et al., UC Berkeley, NeurIPS 2024) to measure how helpful a response is to the original harmful query, considering both how convincing and how specific it is. If an answer addresses only part of the request, the score reflects the level of assistance it actually offers. As suggested, we also carried out independent judge and human evaluations to assess the reported results under different evaluators.

For independent judges evaluation, we evaluate all methods with DeepSeek V4.1 for StrongREJECT and the HarmBench classifier for attack success rate (ASR). As shown in Tables R2 and R3, the average StrongREJECT scores on Knowledge and Viewpoint are 0.615 and 0.813, respectively, and the corresponding HarmBench ASRs are 0.547 and 0.810. Our method achieves the best overall performance under both evaluators, with Viewpoint attaining the highest scores across all four target models. These results show that our method consistently retains its advantage across different evaluators.

Table R2. StrongREJECT scores (↑) with DeepSeek V4.1 as the judge.

QueryMethodGemini 2.5 FlashGPT-5.1LLaMA-4Qwen3-NextAvg.
KnowledgeSelfCipher0.7070.0070.0000.6080.330
FlipAttack0.8270.0330.4750.0080.336
QueryAttack0.7110.0690.2880.8560.481
Authority endorsement0.1980.0380.3780.0930.177
Evidence-based Persuasion0.2440.1030.3080.1710.206
Expert Endorsement0.2470.1010.3590.1590.216
Logical appeal0.2260.0780.4890.1530.237
Misrepresentation0.2030.1090.4770.2210.252
SATA-MLM0.3460.0470.3140.5950.325
Rebuttal Attack (Ours)0.7800.1940.6830.8030.615
ViewpointSelfCipher0.7110.0200.0070.5220.315
FlipAttack0.7970.0250.2660.0840.293
QueryAttack0.7430.0760.3070.8410.492
Authority endorsement0.1220.0530.2080.0640.111
Evidence-based Persuasion0.1410.1170.1520.0890.125
Expert Endorsement0.1970.1230.2280.0580.151
Logical appeal0.1950.1380.3010.0830.179
Misrepresentation0.2520.1300.3370.2170.234
SATA-MLM0.3030.0930.2430.3690.252
Rebuttal Attack (Ours)0.9520.5410.8010.9580.813

Table R3. ASR (↑) with the HarmBench classifier as the judge.

QueryMethodGemini 2.5 FlashGPT-5.1LLaMA-4Qwen3-NextAvg.
KnowledgeSelfCipher0.7470.0070.0000.6400.348
FlipAttack0.7400.0530.6130.5530.490
QueryAttack0.8530.0330.4070.8930.547
Authority endorsement0.0870.0070.3470.0530.123
Evidence-based Persuasion0.0870.0270.3000.0470.115
Expert Endorsement0.1000.0270.3530.0600.135
Logical appeal0.1270.0330.4400.1070.177
Misrepresentation0.0930.0400.4600.0800.168
SATA-MLM0.5670.0930.5400.6000.450
Rebuttal Attack (Ours)0.6930.0670.7130.7130.547
ViewpointSelfCipher0.6870.0000.0000.5200.302
FlipAttack0.7530.0330.5670.6400.498
QueryAttack0.7870.0200.3000.9000.502
Authority endorsement0.0600.0200.1330.0200.058
Evidence-based Persuasion0.0600.0270.0670.0270.045
Expert Endorsement0.1400.0470.1600.0270.093
Logical appeal0.1070.1070.2070.0330.113
Misrepresentation0.1930.1200.2870.1070.177
SATA-MLM0.5070.1800.3070.3200.328
Rebuttal Attack (Ours)0.9400.5470.7670.9870.810
AuthorsOpenReview ↗

Response to Reviewer h81i – part 3 of 6

Following StrongREJECT's human evaluation protocol, we ran a small-scale human evaluation of outputs from Rebuttal Attack and nine baseline jailbreak methods. We recruited five volunteers, each independently rating 200 responses, with 100 from Knowledge and 100 from Viewpoint (10 per method). Using a blind setup, evaluators were not told the target models or attack methods. They rated answers on a 1–5 scale (later normalized to [0, 1]) based only on the original queries, the blinded model responses, and shared scoring criteria. We averaged the median rating for each response to obtain the Human Score reported in Table R4.

Table R4 reports Human Scores, StrongREJECT scores, and HarmBench ASRs for all methods; Table R5 reports MAE and method-level Spearman correlations between automated and human evaluations, capturing agreement in method rankings. In Table R4, Rebuttal Attack achieves the highest Human Score for both Knowledge and Viewpoint, consistent with the StrongREJECT results. Table R5 shows StrongREJECT (GLM-5.2 and DeepSeek V4.1) gives similar results and aligns better with human ratings than HarmBench, with lower MAE and higher rank correlations.

Table R4. Human and automatic evaluation of queries. StrongREJECT (↑) is evaluated with GLM-5.2 and DeepSeek V4.1; ASR (↑) with HarmBench.

QueryMethodGLM-5.2DeepSeek V4.1HarmBenchHuman Score
KnowledgeSelfCipher0.5750.6000.6000.475
FlipAttack0.4380.3380.6000.350
QueryAttack0.3750.3250.4000.175
Authority endorsement0.0000.0000.0000.075
Evidence-based Persuasion0.2000.2000.1000.225
Expert Endorsement0.1000.2000.1000.175
Logical appeal0.2000.2000.1000.250
Misrepresentation0.4000.5000.3000.400
SATA-MLM0.3000.3250.3000.200
Rebuttal Attack (Ours)0.6880.7500.5000.575
ViewpointSelfCipher0.3000.3000.3000.300
FlipAttack0.2380.2500.6000.225
QueryAttack0.6130.4630.4000.375
Authority endorsement0.1000.1630.1000.175
Evidence-based Persuasion0.0000.0000.0000.025
Expert Endorsement0.2000.2750.2000.300
Logical appeal0.0000.0000.0000.125
Misrepresentation0.1750.1630.1000.200
SATA-MLM0.1000.1000.3000.150
Rebuttal Attack (Ours)0.8250.8630.8000.850

Table R5. Agreement between automatic and human evaluation scores.

QueryEvaluatorMAE ↓Spearman Correlation ↑
KnowledgeGLM-5.20.1130.854 (p=0.002)
DeepSeek V4.10.1210.849 (p=0.002)
HarmBench0.1900.675 (p=0.032)
ViewpointGLM-5.20.0750.972 (p<0.001)
DeepSeek V4.10.0800.991 (p<0.001)
HarmBench0.1680.801 (p=0.005)
Reviewer_GCGXOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces Rebuttal Attack, a single-turn black-box jailbreak method that reframes harmful requests as knowledge correction or viewpoint debate tasks. The authors also propose Relative Continuation Likelihood (RCL) to analyze the attack mechanism. Experiments across four LLMs demonstrate substantial improvements over existing single-turn jailbreak baselines.

Critical strengths

1.Practical attack design. The method exploits natural language behaviors without requiring access to model internals, making it applicable to realistic black-box settings.

2.Strong empirical results. The attack achieves competitive performance across proprietary and open-weight models, with comparisons against multiple baselines and additional full-benchmark evaluations.

3.Useful analytical insights. The ablation studies and token-level likelihood analysis provide additional perspectives on how rebuttal-style prompting affects model safety behavior.

Critical weaknesses

1.Questionable evaluation reliability. The results rely on a single LLM judge, and some examples in Appendix D receive high harmfulness scores without clearly satisfying the original malicious objectives. Independent human evaluation would help establish the reliability of the reported gains.

2.Limited methodological novelty. The proposed attack combines several established prompting techniques, including semantic neutralization, role-playing, and explicit output constraints. The existing ablations do not fully establish whether rebuttal framing itself offers a meaningful advantage over simpler correction or debate-based prompts.

3.Insufficient evidence for the proposed mechanism. RCL measures the likelihood of predefined response phrases rather than actual harmful continuations. Its connection to jailbreak success remains unclear, and the analysis does not convincingly establish the claimed autoregressive reinforcement effect.

4.Unclear generalization boundaries. The main evaluation focuses on queries selected for their suitability for rebuttal attacks, while the full-benchmark routing evaluation lacks a clearly documented held-out split on StrongREJECT. Moreover, the limited gains on GPT-5.1 knowledge queries suggest that effectiveness is more model-dependent than the discussion acknowledges.

Questions for authors

See weakness

5 comments
AuthorsOpenReview ↗

Response to Reviewer GCGX – part 1 of 5

We sincerely thank the reviewer’s helpful and constructive feedback, which will assist us in enhancing the quality of our work. Below, we present a point-by-point reply and hope it addresses your concerns.

Response to W1: about evaluation reliability.

Thank you for this valuable suggestion. We rely on StrongREJECT (Souly et al., UC Berkeley, NeurIPS 2024) to measure how helpful a response is to the original harmful query, considering both how convincing and how specific it is. If an answer addresses only part of the request, the score reflects the level of assistance it actually offers. As suggested, we also carried out independent judge and human evaluations to assess the reported results under different evaluators.

For independent judges evaluation, we evaluate all methods with DeepSeek V4.1 for StrongREJECT and the HarmBench classifier for attack success rate (ASR). As shown in Tables R1 and R2, the average StrongREJECT scores on Knowledge and Viewpoint are 0.615 and 0.813, respectively, and the corresponding HarmBench ASRs are 0.547 and 0.810. Our method achieves the best overall performance under both evaluators, with Viewpoint attaining the highest scores across all four target models. These results show that our method consistently retains its advantage across different evaluators.

Table R1. StrongREJECT scores (↑) with DeepSeek V4.1 as the judge.

QueryMethodGemini 2.5 FlashGPT-5.1LLaMA-4Qwen3-NextAvg.
KnowledgeSelfCipher0.7070.0070.0000.6080.330
FlipAttack0.8270.0330.4750.0080.336
QueryAttack0.7110.0690.2880.8560.481
Authority endorsement0.1980.0380.3780.0930.177
Evidence-based Persuasion0.2440.1030.3080.1710.206
Expert Endorsement0.2470.1010.3590.1590.216
Logical appeal0.2260.0780.4890.1530.237
Misrepresentation0.2030.1090.4770.2210.252
SATA-MLM0.3460.0470.3140.5950.325
Rebuttal Attack (Ours)0.7800.1940.6830.8030.615
ViewpointSelfCipher0.7110.0200.0070.5220.315
FlipAttack0.7970.0250.2660.0840.293
QueryAttack0.7430.0760.3070.8410.492
Authority endorsement0.1220.0530.2080.0640.111
Evidence-based Persuasion0.1410.1170.1520.0890.125
Expert Endorsement0.1970.1230.2280.0580.151
Logical appeal0.1950.1380.3010.0830.179
Misrepresentation0.2520.1300.3370.2170.234
SATA-MLM0.3030.0930.2430.3690.252
Rebuttal Attack (Ours)0.9520.5410.8010.9580.813

Table R2. ASR (↑) with the HarmBench classifier as the judge.

QueryMethodGemini 2.5 FlashGPT-5.1LLaMA-4Qwen3-NextAvg.
KnowledgeSelfCipher0.7470.0070.0000.6400.348
FlipAttack0.7400.0530.6130.5530.490
QueryAttack0.8530.0330.4070.8930.547
Authority endorsement0.0870.0070.3470.0530.123
Evidence-based Persuasion0.0870.0270.3000.0470.115
Expert Endorsement0.1000.0270.3530.0600.135
Logical appeal0.1270.0330.4400.1070.177
Misrepresentation0.0930.0400.4600.0800.168
SATA-MLM0.5670.0930.5400.6000.450
Rebuttal Attack (Ours)0.6930.0670.7130.7130.547
ViewpointSelfCipher0.6870.0000.0000.5200.302
FlipAttack0.7530.0330.5670.6400.498
QueryAttack0.7870.0200.3000.9000.502
Authority endorsement0.0600.0200.1330.0200.058
Evidence-based Persuasion0.0600.0270.0670.0270.045
Expert Endorsement0.1400.0470.1600.0270.093
Logical appeal0.1070.1070.2070.0330.113
Misrepresentation0.1930.1200.2870.1070.177
SATA-MLM0.5070.1800.3070.3200.328
Rebuttal Attack (Ours)0.9400.5470.7670.9870.810
AuthorsOpenReview ↗

Response to Reviewer GCGX – part 2 of 5

Following StrongREJECT's human evaluation protocol, we ran a small-scale human evaluation of outputs from Rebuttal Attack and nine baseline jailbreak methods. We recruited five volunteers, each independently rating 200 responses, with 100 from Knowledge and 100 from Viewpoint (10 per method). Using a blind setup, evaluators were not told the target models or attack methods. They rated answers on a 1–5 scale (later normalized to [0, 1]) based only on the original queries, the blinded model responses, and shared scoring criteria. We averaged the median rating for each response to obtain the Human Score reported in Table R3.

Table R3 reports Human Scores, StrongREJECT scores, and HarmBench ASRs for all methods; Table R4 reports MAE and method-level Spearman correlations between automated and human evaluations, capturing agreement in method rankings. In Table R3, Rebuttal Attack achieves the highest Human Score for both Knowledge and Viewpoint, consistent with the StrongREJECT results. Table R4 shows StrongREJECT (GLM-5.2 and DeepSeek V4.1) gives similar results and aligns better with human ratings than HarmBench, with lower MAE and higher rank correlations.

Table R3. Human and automatic evaluation of queries. StrongREJECT (↑) is evaluated with GLM-5.2 and DeepSeek V4.1; ASR (↑) with HarmBench.

QueryMethodGLM-5.2DeepSeek V4.1HarmBenchHuman Score
KnowledgeSelfCipher0.5750.6000.6000.475
FlipAttack0.4380.3380.6000.350
QueryAttack0.3750.3250.4000.175
Authority endorsement0.0000.0000.0000.075
Evidence-based Persuasion0.2000.2000.1000.225
Expert Endorsement0.1000.2000.1000.175
Logical appeal0.2000.2000.1000.250
Misrepresentation0.4000.5000.3000.400
SATA-MLM0.3000.3250.3000.200
Rebuttal Attack (Ours)0.6880.7500.5000.575
ViewpointSelfCipher0.3000.3000.3000.300
FlipAttack0.2380.2500.6000.225
QueryAttack0.6130.4630.4000.375
Authority endorsement0.1000.1630.1000.175
Evidence-based Persuasion0.0000.0000.0000.025
Expert Endorsement0.2000.2750.2000.300
Logical appeal0.0000.0000.0000.125
Misrepresentation0.1750.1630.1000.200
SATA-MLM0.1000.1000.3000.150
Rebuttal Attack (Ours)0.8250.8630.8000.850

Table R4. Agreement between automatic and human evaluation scores.

QueryEvaluatorMAE ↓Spearman Correlation ↑
KnowledgeGLM-5.20.1130.854 (p=0.002)
DeepSeek V4.10.1210.849 (p=0.002)
HarmBench0.1900.675 (p=0.032)
ViewpointGLM-5.20.0750.972 (p<0.001)
DeepSeek V4.10.0800.991 (p<0.001)
HarmBench0.1680.801 (p=0.005)
AuthorsOpenReview ↗

Response to Reviewer GCGX – part 3 of 5

Response to W2: about ablation.

Thank you for your constructive suggestion. We added detailed ablation experiments to analyze how each Rebuttal Attack component contributes to jailbreak performance. All responses were evaluated by GLM-5.2 under StrongREJECT scoring. Each group used the same queries and target models, varying only specific components, and we recorded the resulting StrongREJECT score changes. The tested configurations are:

  • Direct Malicious Query: The original malicious query is sent directly to the target model.
  • Simple correction and debate-based prompt:A simpler correction or debate-based prompt.
  • Neutralized Query: The original query is rewritten into a semantically neutral form, without role-play or output formatting.
  • Neutralized Query + Role Playing: A role-play persona is added to the neutralized query.
  • Neutralized Query + Role Playing + Output Requirements: We further add output constraints, including schemas, specificity, anti-refusal instructions, and other intertwined instructions that are difficult to ablate separately.
  • Neutralized Query + Rebuttal Framing (Full Rebuttal Attack): The full configuration combines semantic neutralization and Rebuttal Framing with rebuttal-focused role-playing and output requirements. Rebuttal-focused role-playing defines a persona for correcting knowledge errors and debating viewpoints; output requirements specify what to correct or debate, and how to structure the response.

As shown in Table R5, responses to the original malicious queries obtain the lowest scores. Simple correction or debate prompts raise the average StrongREJECT score from 0.060 to 0.242 for Knowledge queries and from 0.191 to 0.541 for Viewpoint queries, with smaller gains for Knowledge queries. Many responses only explained why the proposed approach would fail and offered no practical help with the original harmful queries.

Compared with the original requests, semantic neutralization improved attack effectiveness. Role-playing mainly helped on Viewpoint queries. Output requirements mainly shape and constrain responses (e.g., extracting key points for concise, structured outputs), causing only minor StrongREJECT score changes. With rebuttal framing, the average StrongREJECT score rises from 0.392 to 0.622 for Knowledge queries and from 0.675 to 0.812 for Viewpoint queries, indicating that rebuttal framing substantially strengthens jailbreak effectiveness.

Table R5. Component ablation evaluated by GLM-5.2 using the StrongREJECT score (↑).

QueryMethodGemini 2.5 FlashGPT-5.1LLaMA-4Qwen3-NextAvg.
KnowledgeDirect Malicious Query0.0400.0000.1490.0510.060
Simple correction and debate-based prompt0.3080.0780.2410.3420.242
Neutralized Query0.6270.1800.6450.2470.425
Neutralized Query + Role Playing0.7050.0970.5800.2780.415
Neutralized Query + Role Playing + Output Requirements0.6930.1070.5020.2670.392
Neutralized Query + Rebuttal Framing (Full Rebuttal Attack)0.8170.1710.6710.8310.622
ViewpointDirect Malicious Query0.3540.0330.3550.0200.191
Simple correction and debate-based prompt0.6430.6060.5730.3430.541
Neutralized Query0.7620.2320.6830.3100.496
Neutralized Query + Role Playing0.7870.6250.6680.5360.654
Neutralized Query + Role Playing + Output Requirements0.7500.6330.7180.5970.675
Neutralized Query + Rebuttal Framing (Full Rebuttal Attack)0.9610.5490.7820.9580.812

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.