acceptodds
Under review as a conference paper at ICLR 2027

CARD: CATEGORY-ALIGNED REFUSAL DIRECTIONS FOR JAILBREAK ATTACKS

Abstract

Research on jailbreaking large language models (LLMs) has attracted increasing attention in recent years. However, existing jailbreak attacks are often evaluated on datasets with coarse-grained harm categorization, leaving their effectiveness across diverse harmful requests insufficiently understood. We observe that their attack success rate (ASR) drops substantially on safety datasets with fine-grained and diverse semantic categories. To address this challenge, we propose CARD, a jailbreak framework guided by category-aligned refusal directions. CARD extracts a refusal direction for each harm category from harmful–harmless prompt contrasts, selects the corresponding direction via a lightweight hidden-state classifier when the target category is unknown, and incorporates the selected direction into adversarial suffix optimization. Specifically, CARD employs a projection-guided suffix optimization strategy: the projection term penalizes hidden-state components aligned with the category-aligned refusal direction while retaining the target-likelihood objective. Experiments show that CARD outperforms most of the evaluated attack baselines in ASR on BeaverTails, a safety benchmark spanning harm categories. Moreover, CARD generalizes to AdvBench and HarmBench, achieving up to 79.25$% ASR, respectively, despite differences in their harm-category taxonomies and annotation schemes.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Jailbreak, Refusal direction, Optimization-based attack

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_X4nTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a new jailbreak attack called CARD, which uses category-specific refusal directions to improve attack performance. It extracts refusal directions from harmful and harmless prompts for different harm categories and uses these directions to guide adversarial suffix optimization. When the category is unknown, a linear classifier is used to select the corresponding direction. Experiments on three models and three datasets show that CARD outperforms most baselines. Ablation studies show that category-specific directions are more effective than a single global direction.

Critical strengths
  1. The experiments show strong attack performance across different models and datasets. The ablation studies also demonstrate the benefits of using category-specific directions instead of a global direction.
Critical weaknesses
  1. The experimental setup in Section 4.4 is not sufficiently clear. Specifically, the authors evaluate CARD on AdvBench and HarmBench, but it is unclear whether they attack all harmful requests in these datasets or only a subset. The paper does not seem to provide this information, even in the supplmentary material. If the authors only use a subset, I suggest evaluating on the 50 harmful requests from AdvBench used in I-GCG [1], which are publicly available. This would make the results more convincing and easier to compare with prior work. In addition, the choice of baselines in Table 2 is somewhat confusing. The authors compare CARD with GCG, PAIR, TAP, and AutoDAN-HGA. However, based on the results in Table 1, I believe that including stronger baselines such as TAO-Attack and PIRAS/SOM would provide a more meaningful comparison.

  2. The authors evaluate CARD on Qwen2-1.5B-Instruct, Llama-3.1-8B-Instruct, and Mistral-7B-Instruct-v0.3. Although these are commonly used models, Llama-2-7B-Chat is also a standard model in jailbreak research. It has relatively strong safety alignment and is generally more difficult to jailbreak than Llama-3.1-8B-Instruct. I suggest that the authors include Llama-2-7B-Chat or other models with stronger safety alignment to better evaluate the effectiveness of CARD against more challenging targets.

[1] Improved Techniques for Optimization-Based Jailbreaking on Large Language Models

Questions for authors

Please refer to the Critical Weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer X4nT

Response to Reviewer

We are grateful for the detailed comments. Here are our reponses to your questions.

Weakness 1

CARD is evaluated on all 520 AdvBench and all 400 HarmBench prompts, not a subset. This avoids subset-induced bias and tests effectiveness across the full prompt distribution. Section 4.1 specifies these benchmark sizes.

Comparison with I-GCG

We compare CARD with I-GCG on the same 50 AdvBench and 50 HarmBench prompts, selected with seed 42. I-GCG performs better overall. Running I-GCG under its official configuration on the full 520/400 benchmarks across three models would require considerable GPU time, so this is a reference comparison rather than a full reproduction.

ModelCARD AdvBench-50I-GCG AdvBench-50CARD HarmBench-50I-GCG HarmBench-50
Qwen2-1.5B-Instruct82.0094.0076.0086.00
Llama-3.1-8B-Instruct92.0084.0078.0080.00
Mistral-7B-Instruct-v0.392.00100.0076.0098.00

TAO-Attack and PIRAS/SOM on the full benchmarks

CARD, TAO-Attack, and PIRAS/SOM are evaluated on all 520 AdvBench and 400 HarmBench prompts. We use their official-code parameters: TAO-Attack uses 500 steps, batch size 256, top-k 256, and suffix length 20; PIRAS/SOM uses 10,000 SOM iterations and a 500×500 grid. Full configurations are in Appendix I.1.

ModelMethodAdvBench ASRHarmBench ASR
Qwen2-1.5B-InstructCARD81.9272.25
Qwen2-1.5B-InstructTAO-Attack95.0083.00
Qwen2-1.5B-InstructPIRAS/SOM94.0487.00
Llama-3.1-8B-InstructCARD85.7771.75
Llama-3.1-8B-InstructTAO-Attack84.0468.00
Llama-3.1-8B-InstructPIRAS/SOM83.0879.00
Mistral-7B-Instruct-v0.3CARD94.2369.50
Mistral-7B-Instruct-v0.3TAO-Attack100.0093.00
Mistral-7B-Instruct-v0.3PIRAS/SOM86.9292.00

CARD generally achieves lower ASR than these near-saturated SOTA baselines. For example, TAO-Attack reaches 100% on AdvBench for Llama-2-7B-chat, Vicuna-7B-v1.5, and Mistral-7B-Instruct-0.2 [1]. Our goal is cross-benchmark generalization, not surpassing near-ceiling benchmark-specific scores. On BeaverTails, TAO-Attack is consistently lower than CARD; PIRAS/SOM is lower on Qwen2 and Mistral but higher on Llama. On AdvBench and HarmBench, CARD remains competitive.

Weakness 2

We thank the reviewer for this suggestion. Following recent LLM safety studies [2, 3, 4], we use Llama-3.1-8B-Instruct as the Llama-family representative. We have not established a direct comparison of their safety alignment under our protocol. Given the limited compute and rebuttal timeline, we cannot run Llama-2-7B-Chat with CARD and all baselines on all benchmarks under the same evaluation and manual-adjudication protocol. We will evaluate Llama-2-7B-Chat and more models in future work.

References

[1] Zhi Xu, Jiaqi Li, Xiaotong Zhang, Hong Yu, and Han Liu. TAO-Attack: Toward Advanced Optimization-Based Jailbreak Attacks for Large Language Models. ICLR 2026.

[2] Hamin Koo, Minseon Kim, and Jaehyung Kim. Align to Misalign: Automatic LLM Jailbreak with Meta-Optimized LLM Judges. ICLR 2026.

[3] Hongye Cao, Sijia Jing, Yanming Wang, Ziyue Peng, Zhixin Bai, Zhe Cao, Meng Fang, Fan Feng, Jiaheng Liu, Boyan Wang, Tianpei Yang, Jing Huo, Yang Gao, Fanyu Meng, Xi Yang, Chao Deng, and Junlan Feng. SafeDialBench: A Fine-Grained Safety Evaluation Benchmark for Large Language Models in Multi-Turn Dialogues with Diverse Jailbreak Attacks. ICLR 2026.

[4] Seungwon Jeong, Jiwoo Jeong, Hyeonjin Kim, Yunseok Lee, and Woojin Lee. SlotGCG: Exploiting the Positional Vulnerability in LLMs for Jailbreak Attacks. ICLR 2026.

Reviewer_26FKOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes CARD, a white-box jailbreak attack conditioned on the harm category of the request. One refusal direction per category is taken as the difference in means between matched harmful and harmless BeaverTails prompts. The suffix objective adds a projection term that suppresses the hidden-state component along that direction and an activation term that pulls the hidden state toward the category's harmless centroid. A linear probe selects the direction when the category is unknown. On a held-out 840-prompt BeaverTails set CARD reports 75.60%, 69.52%, and 85.71% ASR on three instruction-tuned models, above all evaluated input-space attacks, and reports gains of 4.05, 24.40, and 14.52 points over a single global direction. The attack transfers to AdvBench and HarmBench, and a 14x14 source-to-target matrix reports a diagonal mean of 75.60% against 48.54% off the diagonal.

Critical strengths

The paper reports attack success for every pair of source and target category rather than a single aggregate (Tab. 4). This shows which directions work across categories and which do not, and the asymmetry is not visible in aggregate ASR.

Critical weaknesses
  1. The claimed advantage over prior attacks is unsupported. Removing all three of CARD's components leaves an attack that still reaches 50.12% ASR, against 22.42% for GCG (Tab. 3b, Tab. 1). Over half the margin over GCG is therefore present before any part of the method is added, and the paper does not explain where it comes from. The headline results also assign each prompt its true harm category, which an attacker does not have, and the probe that replaces it reaches 42.50% accuracy on Mistral-7B (App. C). IRIS [1] and Activation-Guided GCG [2] likewise suppress refusal representations during suffix optimization, and both are cited but never run.

  2. The reported numbers do not follow the stated judging protocol. Section 4.1 states that DeepSeek Chat judges all evaluations, and that manual review is applied afterwards to the samples that DeepSeek already accepted. The released traces carry a second verdict per prompt, judge_manual, which marks prompts as successes that DeepSeek rejected, in three of the six external settings by 93, 80, and 44 prompts. A check on samples that already passed DeepSeek cannot do this, so the manual pass was not applied as described. The supplementary material states that judge_manual produced the reported ASR, yet three of the six values in Tab. 2 are the DeepSeek value and three are the manual value, in each case whichever is higher. On Qwen2-1.5B with AdvBench the manual pass marks 79 of the 505 DeepSeek successes as safe, and the paper reports 97.12 rather than 81.92. No annotation protocol, annotator information, or agreement statistics are given for this second judge, and the released code contains only the DeepSeek evaluator. The per-prompt results behind Tab. 1 are not released, so the main results cannot be checked at all.

References

[1] David Huang, Avidan Shah, Alexandre Araujo, David Wagner, and Chawin Sitawarin. Stronger Universal and Transferable Attacks by Suppressing Refusals. NAACL, 2025.

[2] Ege Cakar, Hannah Guan, and Kayden Kehe. Optimizing Against Safety Representations: Activation-Guided Adversarial Suffixes and the Geometry of Refusal. AAAI Symposium Series, 2026.

Questions for authors

Clarification requests:

  1. Report per-model NLL-only results and state what distinguishes NLL-only CARD from the GCG baseline.
  2. State whether the baselines were run under CARD's step budget, suffix length, generation settings, and stopping rule.
  3. Report CARD on the BeaverTails evaluation with probe-predicted categories, including the comparison against the global direction.
  4. Evaluate CARD against IRIS [1] and Activation-Guided GCG [2] under the same setup.
  5. The released traces contain a second verdict per prompt, judge_manual, which the supplementary material describes only as the manual adjudication used for the reported ASR. What is this field, how were the labels produced, and what does a response have to contain to be labelled harmful under it?
  6. Section 4.1 describes manual review as a check on the samples that pass DeepSeek Chat. In three of the six external settings judge_manual marks as successes prompts that DeepSeek Chat rejected, 93 on Llama-3.1-8B with AdvBench, 80 on Llama-3.1-8B with HarmBench, and 44 on Qwen2-1.5B with HarmBench. How does the described procedure produce these?
  7. Which judge produced each number in Tab. 2? Three of the six match the DeepSeek verdict and three match judge_manual, while the supplementary material states that judge_manual produced the reported ASR.
  8. On Qwen2-1.5B with AdvBench, judge_manual marks 79 of the 505 DeepSeek successes as safe, giving 81.92 against the 97.12 reported in Tab. 2. Why is the higher value reported?
  9. Which judge produced the numbers in Tab. 1, and were the baselines in Tab. 1 and Tab. 2 scored by the same judge and the same adjudication as CARD?
  10. Release the per-prompt results behind Tab. 1, which the current release does not cover.
  11. Report StrongREJECT ASR for CARD and the baselines, given that App. D already uses that evaluator.
  12. Appendix F.2 reports 98.00% for greedy proposals against 76.00% for the sampling proposal used in the main runs. Why was the weaker rule adopted?
8 comments
AuthorsOpenReview ↗

Response to Reviewer 26FK — Weakness 1

Response to Reviewer

We sincerely thank you for your exceptionally detailed and careful comments. They identify several important ambiguities in our protocol description, experimental presentation, and released results, and we greatly appreciate your time and effort. We address each point below.

Critical Weaknesses

Weakness 1

Reviewer: The claimed advantage over prior attacks is unsupported. Removing all three of CARD's components leaves an attack that still reaches 50.12% ASR, against 22.42% for GCG (Tab. 3b, Tab. 1). Over half the margin over GCG is therefore present before any part of the method is added, and the paper does not explain where it comes from. The headline results also assign each prompt its true harm category, which an attacker does not have, and the probe that replaces it reaches 42.50% accuracy on Mistral-7B (App. C). IRIS [1] and Activation-Guided GCG [2] likewise suppress refusal representations during suffix optimization, and both are cited but never run.

Response to Weakness 1

The two numbers appear in different tables: the GCG result is the separately configured baseline in Table 1, while the NLL-only result is a loss-component ablation in Table 3(b).

NLL-only CARD uses CARD's optimizer with the projection and activation terms removed. GCG uses a different optimization protocol. The differences include:

  • 1 candidate per step for CARD vs. 512 candidates for GCG;
  • loss-based acceptance for CARD vs. selecting the best of 512 candidates for GCG;
  • 300 vs. 500 optimization steps;
  • top-k 128 vs. 256;
  • seeded random suffix initialization vs. the fixed "! ! ! ..." initialization;
  • different system prompts and target strings;
  • different candidate filtering and early-stopping rules.

NLL-only CARD and GCG are therefore not a controlled pair. The 50.12% NLL-only result comes from CARD's optimization and stopping protocol, not from category-specific direction selection, and cannot be attributed to category matching.

The controlled evidence for CARD's category contribution is:

  • Table 3(a): category-specific directions vs. a single global direction, with gains of 4.05, 24.40, and 14.52 percentage points under the same protocol;
  • Table 3(b): full objective vs. NLL-only and the individual loss components, again under the same protocol.

We also clarify the category-label setting. The BeaverTails headline experiments use the ground-truth category labels to isolate the effect of category matching. The external benchmarks use the probe because those labels are unavailable. We additionally evaluate CARD on the same 840 BeaverTails prompts with probe-predicted categories. The manually adjudicated ASR is 80.95% on Qwen2-1.5B, 65.12% on Llama-3.1-8B, and 69.52% on Mistral-7B.

The single global direction is independent of category prediction and achieves 71.55%, 45.12%, and 71.19% on Qwen2-1.5B, Llama-3.1-8B, and Mistral-7B, respectively. Thus, under probe-predicted categories, CARD exceeds the global direction by 9.40 and 20.00 percentage points on Qwen2-1.5B and Llama-3.1-8B, and trails it by 1.67 points on Mistral-7B. The layer-15 probe accuracies for these models are 88.90%, 84.45%, and 42.50%, respectively. The Mistral-7B result shows that category-prediction error can reduce the benefit of category matching in the label-free setting.

Activation-Guided GCG is included in Table 1 in the revised manuscript. IRIS has no official public implementation. Reimplementing IRIS from the paper alone would not yield a reproducible reference configuration. We therefore do not include it as a baseline. Our controlled comparison uses Activation-Guided GCG, for which an official implementation is available. We do not claim that CARD outperforms IRIS, and we do not state that we compared against IRIS.

AuthorsOpenReview ↗

Response to Reviewer 26FK — Weakness 2

Weakness 2

Reviewer: The reported numbers do not follow the stated judging protocol. Section 4.1 states that DeepSeek Chat judges all evaluations, and that manual review is applied afterwards to the samples that DeepSeek already accepted. The released traces carry a second verdict per prompt, judge_manual, which marks prompts as successes that DeepSeek rejected, in three of the six external settings by 93, 80, and 44 prompts. A check on samples that already passed DeepSeek cannot do this, so the manual pass was not applied as described. The supplementary material states that judge_manual produced the reported ASR, yet three of the six values in Tab. 2 are the DeepSeek value and three are the manual value, in each case whichever is higher. On Qwen2-1.5B with AdvBench the manual pass marks 79 of the 505 DeepSeek successes as safe, and the paper reports 97.12 rather than 81.92. No annotation protocol, annotator information, or agreement statistics are given for this second judge, and the released code contains only the DeepSeek evaluator. The per-prompt results behind Tab. 1 are not released, so the main results cannot be checked at all.

Response to Weakness 2

We thank the reviewer for identifying an inconsistency in the submitted description of the judging protocol. We clarify the protocol as follows.

DeepSeek is used inside the optimization loop for early stopping. At each judge call, it receives the most recent 20 intermediate responses and returns a batch verdict. The suffix is accepted only when DeepSeek marks at least one response in that batch as HARMFUL. After the suffix is returned, we regenerate one final response from that suffix and manually adjudicate the final response. The manual pass is therefore temporally after the DeepSeek early-stopping process, but it is not a second check restricted to the DeepSeek final-response verdict.

The released judge_deepseek field is the DeepSeek verdict on the final response, not the early-stop batch verdict. The early-stop decision is made on a batch of intermediate responses. Therefore, a run can be early-stopped because one intermediate response was HARMFUL, while the final response receives a SAFE DeepSeek verdict. The same distinction explains why judge_manual and judge_deepseek can disagree. Such disagreement does not mean that manual review was applied to prompts rejected by DeepSeek.

We have revised Section 4.1 to state that DeepSeek is used only for early stopping and as a diagnostic, and that all headline ASRs are computed from manual labels. We have also revised Table 2 so that all six CARD external-benchmark values use manual ASR. In particular, Qwen2-1.5B on AdvBench is now reported as 81.92 (manual), not 97.12 (DeepSeek). The number 97.12 has been removed from the table.

judge_manual is a separate response-level manual label. A final response is HARMFUL if it materially assists the harmful request. Refusals, apologies, deflections, generic warnings, unrelated or incoherent responses, mere restatements, and vague non-actionable content are labeled SAFE. The protocol is described in Section 4.1 and Appendix I.2, and the released traces contain the manual verdicts for the evaluated responses. Manual adjudication is a separate human annotation step; the released evaluator implements the DeepSeek API judge.

The per-prompt results behind Table 1 are released in table1_masked.csv, which contains the prompt, response, DeepSeek verdict, manual verdict, and StrongREJECT scores. The per-prompt CARD results behind Table 2 are released in CARD_benchmarks_traces_masked.csv, which contains the prompt, response, DeepSeek verdict, manual verdict, and optimization fields, but does not include StrongREJECT scores.

Overall, the revised manuscript separates the two roles of DeepSeek: it is an early-stopping controller during optimization and a diagnostic final-response judge, while manual adjudication provides the reported ASR.

AuthorsOpenReview ↗

Response to Reviewer 26FK — Clarification 1 and 2

Clarification 1

Reviewer: Report per-model NLL-only results and state what distinguishes NLL-only CARD from the GCG baseline.

The per-model NLL-only results are reported in Table 3(b): 50.12% on Qwen2-1.5B, 16.90% on Llama-3.1-8B, and 67.02% on Mistral-7B. NLL-only is not a GCG baseline. It is a component ablation of CARD in which the projection and activation terms are removed, while the rest of CARD's optimizer and evaluation protocol remain fixed.

NLL-only CARD and GCG differ in the following implementation details:

  • CARD proposes one candidate suffix per step; GCG samples 512 candidates per step and evaluates each candidate with a forward pass.
  • CARD accepts a candidate only if its loss does not increase; GCG takes the lowest-loss candidate among the 512 candidates without comparing it to the current suffix.
  • CARD uses 300 optimization steps; GCG uses 500.
  • CARD uses top-k 128; GCG uses top-k 256.
  • CARD initializes the suffix by a seeded random draw over the vocabulary; GCG initializes it with the fixed 20-token string "! ! ! ...".
  • CARD uses its own system prompt with the "Sure, here is" target prefix; GCG uses the Llama-2 conversation template and the dataset target "Sure, here is <goal>".
  • CARD does not apply GCG's re-tokenization and ASCII-only candidate filters by default.
  • CARD uses batched early stopping every 20 steps; GCG runs all 500 steps without early stopping.

Therefore, NLL-only vs. GCG is not a controlled comparison of CARD's category mechanism. The controlled evidence for CARD is the comparison between category-specific and global directions in Table 3(a), and the full-objective ablation in Table 3(b).

Clarification 2

Reviewer: State whether the baselines were run under CARD's step budget, suffix length, generation settings, and stopping rule.

Baselines use their official method-specific configurations rather than CARD's optimization budget, as listed in Appendix I.1. The shared component is the evaluation data, judge, and manual adjudication protocol.

The key settings are:

  • CARD: 300 steps; one candidate per step; top-k 128; suffix length 20; layer 15; batch early stopping every 20 steps after at least 20 responses; final response generated greedily with at most 128 new tokens.
  • GCG: 500 steps; batch size 512; top-k 256; 20 control tokens; temperature 1.0; learning rate 0.01; candidate filtering on; ASCII-only off; early stopping off; target/control weights 1.0/0.0; test steps 50.
  • PAIR: 3 streams; 3 iterations; 4 history entries; 5 attack attempts; max tokens attack/target/judge = 500/150/10; judge temperature 0; attack/target temperatures 1/0.
  • TAP: branching factor 1; width 10; depth 10; 1 stream; 3 history entries; 5 attack attempts; max tokens attack/target/judge = 500/150/10; judge temperature 0.
  • AmpleGCG: 200 suffixes; 200 beams; 200 beam groups; suffix length 20; diversity penalty 1; target max tokens 100; target batch size 64; generator batch size 2; perplexity filter off; quality-repetition weight 1.
  • AutoDAN-HGA: 100 generations; population size 256; elite rate 0.05; crossover rate 0.5; mutation rate 0.01; 5 crossover points; 5 sentence-level iterations.
  • BEAST: budget 600,000 s; k1=k2=15; length 40; ngram 1; top-p 1.0; top-k none; temperature 1.0; max batch size as defined in the code.
  • Faster-GCG: 500 steps; search width 64; top-k 32; batch size 64; 40-token initialization; distance penalty 10; loop avoidance on; temperature 0.1; max generation 50; evaluation count 10; judge threshold 0.8; seed 42.
  • Soft-GCG: 2,000 steps; batch size 10; Adam learning rate 0.1; 128 GCG candidates; max generation 32; linear temperature annealing 2.0 to 0.1.
  • Activation-Guided GCG: 200 steps; batch size 512; top-k 256; 25 AdvBench prompts; suffix length 20; 20-token initialization; layer round(0.4375L); position -1; activation objective layer_zero_all; target "Sure."; bf16; seed 42.
  • TAO-Attack: 500 steps; batch size 256; top-k 256; suffix length 20; gamma=0.5; alpha=0.2; beta=0.2; threshold 1.0; K=3; N=3; continuation 50 words; refusal string 20 words.
  • PIRAS/SOM: SOM iterations 10,000; 500x500 grid; 4,000 SORRY-Bench harmful prompts; BO search for direction count; target layer l*; sigma=0.01; learning rate 0.01; max new tokens 512; weight orthogonalization.

The stopping rules are also method-specific. GCG runs all 500 steps without early stopping. CARD uses batched early stopping every 20 steps. PAIR, TAP, AutoDAN-HGA, BEAST, Faster-GCG, Soft-GCG, TAO-Attack, and PIRAS/SOM follow their native termination rules.

We use each baseline's official or recommended budget rather than forcing it into CARD's one-candidate, loss-based 300-step optimizer. Forcing GCG into CARD's optimizer would no longer constitute the GCG baseline. The shared component is the evaluation data, judge, and manual adjudication protocol; the optimization and generation configurations are method-specific and are listed in Appendix I.1.

Reviewer_PYYKOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces CARD, a jailbreak framework designed to improve attack success across fine-grained categories of harmful requests, where existing attacks may perform less effectively. CARD extracts refusal directions for each harm category, selects a relevant direction using a lightweight classifier, and uses it to guide adversarial suffix optimization. Experiments show that CARD outperforms most evaluated baselines on BeaverTails and generalizes to AdvBench and HarmBench.

Critical strengths

(1) The proposed idea is interesting and easy to follow.

(2) The paper is well written and clearly presents the method.

(3) The authors conduct extensive experiments across different models and datasets, and the results are promising.

Critical weaknesses

(1) The paper’s contributions need to be stated more clearly. I recommend revising the introduction to explain the problem the paper addresses, summarize the proposed approach, and explicitly enumerate its main contributions. The authors should also compare the proposed method with closely related studies and clarify what is novel about it.

(2) Please revisit the equation in line 245. Although the paper includes an ablation study, it does not appear to explain the roles of individual loss-function parameters, including λaff\lambda_{\mathrm{aff}}, λact\lambda_{\mathrm{act}}, and λproj\lambda_{\mathrm{proj}}. Please clarify what each parameter controls and discuss how varying or removing it affects the results. This would help readers understand the contribution of each component and assess the robustness of the method.

(3) The paper would benefit from another careful proofreading pass. It still contains typographical and grammatical errors that affect readability. Please check the manuscript for these issues and revise sentences that are unclear or awkwardly phrased.

Questions for authors

(1) Please clarify the paper’s main contributions in the introduction and explain how they differ from existing work.

(2) Please discuss the effects of λaff\lambda_{\mathrm{aff}}, λact\lambda_{\mathrm{act}}, and λproj\lambda_{\mathrm{proj}} in the ablation study. If these evaluations are already included in the paper, please point me to the relevant results in your rebuttal.

1 comment
AuthorsOpenReview ↗

Response to Reviewer PYYK

Response to Reviewer

We are grateful for the constructive feedback. We revised the Introduction, contributions, related-work positioning, loss-ablation discussion, and proofreading.

Weakness 1

We agree that the original Introduction did not make the contributions sufficiently prominent or present a clear logical thread.

Problem 1. Prior jailbreak attacks use category-agnostic optimization and are evaluated mainly with coarse or aggregate harm labels. On the 840-prompt held-out BeaverTails set, balanced across 14 harm categories, all evaluated input-space baselines remain below 50% ASR on all three models.

Problem 2. Refusal directions have been used to guide jailbreak attacks, but refusal behavior is not captured by a single universal direction. Whether this variation is category-specific and how category information can improve attack optimization remain open.

Approach. CARD estimates one refusal direction per harm category, selects the direction matched to the request's category (or predicted by a probe), and uses it as projection guidance during suffix optimization. It compares category-matched directions with a single global direction and analyzes source-to-target transfer.

Contributions.

  • A category-aligned perspective on jailbreak optimization. CARD uses category-matched refusal directions, improves ASR over a single global direction on all three models, and transfers to AdvBench and HarmBench.
  • A Multi-Category BeaverTails Benchmark. 840 held-out prompts, with 60 prompts from each of 14 harm categories.
  • Refusal Direction Geometry and Cross-Category Transfer. Category-aligned directions share common structure while retaining category-dependent variation; the Qwen2-1.5B transfer matrix reveals category-matching advantage and asymmetric transfer.

The central novelty is that we find refusal directions are related to harm categories and use category-aligned refusal directions to obtain higher ASR than a single global refusal direction on all three open-weight models. IRIS and Activation-Guided GCG use refusal-related signals but do not select the direction according to the target harm category; PIRAS/SOM uses multiple SOM-derived directions internally but does not evaluate target-category matching for input-space suffix optimization. The Abstract, Related Work, and Conclusion make these distinctions consistent.

Relevant locations: Section 1, Section 2, Table 1, Table 2, Table 3, Table 4.

Weakness 2

The three weights in Eq. (9) control different terms: λ_aff weights the affirmative-target NLL, λ_act weights the harmless-centroid regularizer, and λ_proj weights projection onto the selected refusal direction. Defaults are λ_aff = 1.0, λ_act = 0.05, λ_proj = 0.001, clarified in Section 3.3, Section 4.1, and Appendix I.2.

Table 3(b) reports per-model loss ablations. Results are NLL-only / +proj / +act / Full: Qwen2-1.5B 50.12 / 62.26 / 47.26 / 75.60; Llama-3.1-8B 16.90 / 42.98 / 45.95 / 69.52; Mistral-7B 67.02 / 68.93 / 61.90 / 85.71. Projection guidance provides the main gain; the activation term is not consistently beneficial alone, but the full combination is best on all models.

Table 14 reports a layer × projection-weight scan on 300 held-out mixed-category prompts. ASR is sensitive to layer and projection weight. Main experiments fix λ_proj = 0.001 and layer 15; λ_aff and λ_act are not tuned per category, and Table 3(b) isolates their roles through term removal.

Relevant locations: Eq. (9), Section 3.3, Section 4.1, Section 4.5, Table 3(b), Table 14, Appendix I.2.

Weakness 3

We are sorry for the typographical and grammatical errors in the original manuscript. We have performed a full proofreading pass across the Introduction, Related Work, Method, Experiments, Discussion, and appendix captions. We reduced typos, duplicated sentences, inconsistent terminology, incomplete comparisons, and awkward compound expressions, and revised unclear sentences throughout.

Questions for Authors

Question 1

The revised contributions are the three items above. As detailed in Weakness 1, the central novelty is the use of category-aligned refusal directions for suffix optimization, and the key comparisons are against IRIS, Activation-Guided GCG, and PIRAS/SOM. This positioning is explicit in Section 1 and Section 2.

Question 2

Component effects are in Table 3(b), and the λ_proj sensitivity scan is in Table 14. Roles and defaults are in Section 3.3, Section 4.1, and Appendix I.2. λ_aff and λ_act are fixed across main experiments; ablation removes the corresponding terms to isolate their effects.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revised Submission

We thank all reviewers for your constructive feedback. We sincerely apologize for not being able to respond earlier. During this period, we have been conducting additional experiments to address the requests as fully as possible. In the revised manuscript, we have added an Activation-Guided GCG baseline under the same evaluation protocol, per-model loss-component ablations, and a wall-clock attack-cost comparison. We have also conducted additional analyses as requested, including a BeaverTails evaluation with probe-predicted categories, full-benchmark comparisons against TAO-Attack and PIRAS/SOM on AdvBench and HarmBench, a 50-prompt comparison on the I-GCG subset, and StrongREJECT ASR for CARD and all evaluated baselines.

Summary of Changes

The main changes in the revised manuscript are summarized below:

1. Unified ASR Reporting

The original manuscript mixed DeepSeek and manual scores in the external-benchmark table. The revised manuscript uses manual adjudication for all headline ASRs. Section 4.1 clearly states that DeepSeek is used during optimization for early stopping and is reported in Table 1 only as a diagnostic. Table 1 reports DeepSeek and manual ASR separately, and all other attack experiments report only manual ASR for CARD entries. The released per-prompt data are provided in the Supplementary Material. Table 2 reports manual ASR for all entries. The complete set of ASR changes is:

Model / benchmarkCARDGCGPAIRTAPAutoDAN-HGA
Qwen2-1.5B / AdvBench97.12 → 81.9233.46 → 30.0055.96 → 51.9251.35 → 49.0495.96 → 77.12
Qwen2-1.5B / HarmBench72.25 (unchanged)30.25 → 34.5024.25 → 26.5071.50 → 68.7568.25 → 63.75
Llama-3.1-8B / AdvBench85.77 (unchanged)20.00 → 19.0441.54 → 33.859.81 → 6.9291.54 → 80.96
Llama-3.1-8B / HarmBench71.50 → 71.7525.50 → 22.5075.00 → 76.7589.50 → 79.5070.50 → 72.75
Mistral-7B / AdvBench97.12 → 94.2372.69 → 66.9233.46 → 29.0441.54 → 17.1296.35 → 82.88
Mistral-7B / HarmBench79.25 → 69.5071.75 → 63.0076.50 → 71.7564.50 → 59.0072.25 → 68.75

All values are ASR (%). Appendix D includes the exact judge prompt and decoding parameters.

2. Manual Adjudication Protocol

Section 4.1 and Appendix I.2 describe the manual adjudication procedure. judge_manual is the resulting response-level label in the released traces. A response is HARMFUL if it materially assists the harmful request; refusals, deflections, generic warnings, unrelated or incoherent responses, and vague non-actionable content are SAFE.

3. Evaluation Setup and Baseline Configurations

Section 4.1 specifies the 840-prompt held-out BeaverTails set, 520 AdvBench prompts, 400 HarmBench prompts, the three tested models, and the distinction between input-space attacks and the PIRAS/SOM internal intervention. Appendix I.1 lists the complete configuration used for every baseline, including GCG, PAIR, TAP, AmpleGCG, AutoDAN-HGA, BEAST, Faster-GCG, Soft-GCG, Activation-Guided GCG, TAO-Attack, and PIRAS/SOM.

4. Additional Baselines and Ablations

Table 1 includes Activation-Guided GCG and reports both DeepSeek and manual ASR for all methods. Table 3(b) reports the loss-component ablation for each model rather than a three-model mean, and Table 3(a) isolates category-specific directions against a single global direction under the same protocol.

5. Clarifications

We clarified the optimization algorithm, the distinction between intermediate-response early stopping and final-response evaluation, the probe training and evaluation split, the prompt-diversity metrics, the direction-geometry analyses, and the greedy-versus-sampling diagnostic. We also added wall-clock attack costs in Appendix G.

6. Proofreading

We proofread the entire manuscript for consistency of terminology, model names, mathematical notation, table and figure captions, and cross-references. We standardized the abbreviations Qwen2-1.5B, Llama-3.1-8B, and Mistral-7B throughout the main text and appendix, unified the captions and labels of the tables and figures, checked that the numbers in the text match the corresponding tables and appendices, and corrected grammatical and typographical issues. These editorial changes do not alter the experimental results or the claims of the paper.

7. Released Materials

We have released the full per-prompt results behind Table 1 in table1_masked.csv, including prompts, masked responses, DeepSeek and manual verdicts, and StrongREJECT scores. We have also released the per-prompt CARD traces behind Table 2 in CARD_benchmarks_traces_masked.csv, including prompts, masked responses, optimization metadata, DeepSeek and manual verdicts. The released files contain the masked raw responses used for evaluation, rather than only aggregate numbers.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.