acceptodds
Under review as a conference paper at ICLR 2027

COMBAT: Cross-Model Jailbreaking via Depth-Interpolated Activation Transfer

20 views

Abstract

Extracting intervention directions from the target model is a natural approach to representation-level jailbreaking. Target-native intervention directions already lie in the target representation space, giving them a natural compatibility advantage. We investigate whether refusal and harm-detection directions transferred from another model through representation alignment can nevertheless produce interventions that are as effective as, or even more effective than, target-native ones under the same search budget. We propose rss-odel Jailreaking via Depth-Interpolated ctivation ransfer (), a layerwise representation alignment framework that constructs banks of refusal and harm-detection directions in a source model, maps them into a target hidden space through depth-interpolated bridges that accommodate unequal model depths, identifies behaviorally effective configurations via constrained two-phase selection, and deploys them through localized interventions with a triangular layer taper. We evaluate 20 ordered source–target pairs from five aligned language models. Held-out state reconstruction assesses the learned cross-model correspondence, while equal-budget target-native controls evaluate the effectiveness of cross-model direction construction. Across four primary benchmarks, Full COMBAT achieves a micro-average GPT-judged attack success rate of 96.2%, compared with 88.7% for the equal-budget target-native control and 20.3% for unmodified targets. Further analyses show that refusal and harm-detection directions make complementary contributions, with the most effective transfers from narrow source-layer bands and trailing source direction-extraction positions immediately preceding assistant generation. The bridges fit held-out states well, and their normalized nonnegative bridge-cosine profiles decay away from selected anchors. Cross-model representation alignment thus informs the construction of effective target-side jailbreak interventions.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Jailbreaking Attack, Cross-Model Transferability, Activation Steering, Representation Alignment, Mechanistic Interpretability

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_7xp2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents COMBAT for transferring refusal and harm-detection directions across aligned LLMs. It extracts directions from a source model, maps them to a target model through ridge-regression bridges, and uses CSTU to select intervention configurations while limiting changes to harmless-prompt behavior. The evaluation covers 20 ordered pairs of five 7–8B models across six safety benchmarks. COMBAT achieves a reported 96.2% ASR, compared with 88.7% for target-native directions under the same configured search budget. The paper also includes ablations and analyses of transferability, bridge quality, and intervention sensitivity.

Critical strengths

The paper addresses an interesting question about whether safety-related directions can transfer across models. I particularly appreciate the motivation drawn from the Platonic Representation Hypothesis, which provides an intuitive and compelling basis for exploring cross-model transferability. The evaluation is comprehensive, covering all 20 source–target pairs. COMBAT consistently outperforms target-native directions under the same configured budget, suggesting potential benefits from cross-model transfer. The randomized bridge and direction controls further support the role of learned cross-model correspondence rather than arbitrary activation perturbations.

Critical weaknesses
  • The practical motivation for white-box cross-model transfer is unclear. COMBAT requires access to the target model's internal representations and activations. Given such access, an attacker could directly optimize attacks against the target rather than transfer directions from another model. Moreover, COMBAT cannot directly support white-box-to-black-box transfer, where transferability would be more practically relevant. This raises concerns about the significance of the proposed threat model.

  • Several baselines consider white-box-to-black-box transfer, where attackers have no internal access to the target model. In contrast, COMBAT requires white-box access to both models, including target-side activation intervention. These methods operate under substantially different threat models and attacker capabilities, making direct ASR comparisons unfair. The reported superiority over these baselines therefore does not convincingly establish COMBAT's advantage.

  • The evaluation is limited to white-box interventions on 7–8B open-weight models. While this is a reasonable setting for studying activation-level attacks, it remains unclear how well the findings generalize to larger models or different alignment procedures. The target models also differ considerably in their initial vulnerability, making aggregate ASR less informative without considering individual targets.

  • The paper demonstrates that transferred directions can weaken safety behavior, but offers limited discussion or evidence of how these findings could inform defensive auditing or mitigation. This mainly limits the scope and significance of the contribution.

  • The comparison with target-native directions is not fully convincing from a computational-cost perspective. The matched budget appears to cover configuration search, while COMBAT additionally requires source-side direction extraction and bridge fitting. Although some of these costs can be amortized, the current results do not establish an advantage under equal end-to-end compute.

Questions for authors
  • What practical scenario motivates white-box cross-model transfer when the attacker can directly attack the target model? What advantage does COMBAT offer over target-specific activation attacks?
  • Several baselines assume black-box target access, while COMBAT requires white-box access. How do the authors justify comparing attack success rates under different attacker capabilities?
  • Does the equal-budget comparison account for direction extraction and bridge fitting, or only configuration search? Would COMBAT retain its advantage under equal end-to-end computation?
  • Why do target-native directions perform substantially worse than mapped directions in the anchor-matched analysis (Appendix E.1)? Is this due to configuration constraints, direction scaling, or intrinsic differences?
  • Does harmless-prompt next-token KL reliably reflect downstream utility? Can the authors provide generation-level or benign-task evaluations to support this claim?
  • To what extent do the findings extend beyond 7–8B open-weight models, particularly to larger models and different alignment procedures?
1 comment
AuthorsOpenReview ↗

Response to Reviewer 7xp2

We agree that target access, practical auditing value and transfer beyond the original scale should be stated more explicitly.

Q1. Target access. COMBAT audits open-weight targets with white-box access. The source is another open-weight model supplying refusal/harm-detection directions; target-layer-specific ridge bridges map them into the target space. Full COMBAT uses both mapped directions. Target Native extracts both directions directly from the target and independently selects its intervention configuration. It is the primary access-matched comparator; black-box input-space attacks are contextual references under different target-access assumptions.

Q2. Auditing value. ASR-GPT counts a response as successful when the GPT judge finds harmful content matching the requested objective. Across 18,925 paired audits with identical benchmark items and prompt identifiers, COMBAT has 17,141 successes versus 16,943 for equal-budget Target Native. Both succeed on 15,985 prompts; COMBAT succeeds on 1,156 prompts where Target Native does not. COMBAT-only successes account for 6.4% of all prompts on which either method succeeds. These 1,156 prompt identifiers and fixed settings form a target-specific regression set for checking whether the same successful jailbreaks persist after mitigation or safety fine-tuning.

Q3. Forward budget. COMBAT preprocessing uses 1,536 source + 512 target model–prompt forwards and 32–48 closed-form ridge fits, with no LLM backward pass. Raw Target Native uses 1,024 target forwards; forward-budget-matched Native gets 1,024 extra target forwards as 16 development-only candidates on the same 64 prompts. Preprocessing plus proxy search totals 102,400–604,160 forwards for each method. Each confirms its own top 96 candidates with matched confirmation counts, target-generation budgets and test sets. COMBAT reaches 88.1% ASR-GPT versus 83.6% (+4.5 pp).

Q4. Direction quality. To separate direction quality from deployment choices and intervention strength, the control fixes the same central target layer, seven-layer triangular window, three intervention sites and α/β strengths for both methods; only mapped versus target-extracted refusal/harm-detection directions change. Across 25 paired runs (five pairs × five seeds), mapped directions improve mean ASR-GPT by +2.5 pp, win on 80.0% of pairs and improve the mean on all five benchmarks. A separate target-native self-ridge control applies ridge filtering using target information only: Full COMBAT reaches 91.7% versus 89.0% ASR-GPT (+2.7 pp), winning on 80.0% of pairs.

Gains vary by target: +5.6 pp for Llama→Qwen, +4.7 pp for Qwen→InternLM and +3.3 pp for Mistral→Llama. Two leave-one-out analyses select methods without the held-out result. Holding out one seed, selection between COMBAT and Target Native uses the other four seeds for that pair and beats always choosing either method. Holding out one benchmark, source-model selection uses the other four benchmarks, adding 16 successes (+0.4 pp) over the paper's fixed cyclic source assignment.

Q5. Benign utility. Against independently selected equal-budget Target Native, Full COMBAT's harmless-prompt KL is 1.6% lower. Downstream tasks complement next-token KL's local distribution measure; the five-target results are:

TaskFull COMBATNativeDifference
GSM8K exact match46.9%45.9%+1.0 pp
ARC-Challenge accuracy74.3%74.7%-0.4 pp
SQuAD F131.2%29.7%+1.5 pp
Pair/task-equal mean50.8%50.1%+0.7 pp

Q6. Cross-size transfer. For Qwen-0.6B (28L)→Vicuna-13B (40L), depth interpolation gives ASR-GPT of 96.0/89.4/88.0/86.0/87.4% on AdvBench/HarmBench/JailbreakBench/MaliciousInstruct/WildJailbreak-AH, respectively. The benign-task results above are a separate five-target evaluation.

The added Qwen3 transfers use 0.6B/1.7B sources (28L) and 4B/8B targets (36L). Interpolation mixes the source layers adjacent to the target's relative-depth position; same index uses the same absolute layer; Hard relative-depth assignment rounds to the nearest source layer. Each cell below lists ASR-GPT (%) in the same five-benchmark order.

Qwen3 source→targetInterpolationSame indexHard relative-depth assignment
0.6B→4B97.0/95.0/94.0/99.0/89.995.0/93.1/93.0/99.0/88.996.0/93.1/95.0/98.0/87.4
0.6B→8B91.5/91.0/92.0/98.0/90.989.9/92.0/90.0/96.0/87.980.9/87.9/85.0/91.0/84.4
1.7B→4B97.0/95.0/97.0/100.0/89.994.5/89.9/95.0/99.0/86.496.5/93.1/97.0/99.0/87.9
1.7B→8B92.0/89.9/90.0/96.0/90.989.4/87.9/88.0/97.0/88.980.4/85.9/78.0/95.0/89.9
20-cell equal mean93.892.090.1

Interpolation ranks first among these rules, supporting the paper default in the evaluated cross-scale settings.

Reviewer_Mv9tOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates cross-model safety representation transfer for activation-based jailbreaking attacks on large language models. The authors propose COMBAT, which extracts refusal and harm-detection directions from a source model and transfers them to a target model through depth interpolation and ridge regression. The framework further employs a two-stage configuration selection strategy and localized activation interventions with triangular layer weighting. Experiments across 20 directed model pairs involving five LLMs demonstrate an average attack success rate of 96.2%, outperforming the target-native baseline of 88.7% under an equal configuration-search budget. Additional ablation studies and representation analyses examine the contributions of depth interpolation, dual-direction interventions, and layer-position selection.

Critical strengths
  1. The paper explores cross-model safety representation transfer and presents an interesting finding that transferred directions can outperform target-native directions, offering valuable insights into shared safety-related representations across LLMs.
  2. The proposed method achieves consistently strong attack performance across 20 directed model pairs and multiple safety benchmarks, demonstrating substantial improvements over the target-native baseline under an equal configuration-search budget.
  3. The paper provides comprehensive ablation studies and additional analyses covering depth interpolation, dual-direction interventions, intervention scope, and randomized mapping controls, which offer empirical support for the effectiveness of the proposed components.
Critical weaknesses
  1. COMBAT primarily integrates existing techniques, including safety direction extraction, linear representation mapping, depth interpolation, and activation steering. The individual components offer relatively limited methodological novelty, and the lack of comparisons with stronger cross-model representation alignment methods makes it difficult to establish the distinct contribution of the proposed framework.
  2. The proposed denoising interpretation of cross-model mapping is primarily supported by a single model-pair analysis. Furthermore, the Task-Agnostic Bridge control differs from Full COMBAT in its training data and configuration selection, while some mechanism analyses reuse intervention strengths selected for COMBAT. These differences make it difficult to isolate the specific contribution of cross-model mapping.
  3. COMBAT and some input-space attack baselines operate under different target-model access assumptions, and a comparison of end-to-end computational costs is lacking. Moreover, the preservation of benign behavior is primarily assessed using next-token KL divergence, without a systematic evaluation of generation quality or performance on benign tasks.
Questions for authors
  1. Could the authors provide comparisons with stronger cross-model representation alignment methods to better establish the advantages of the proposed depth-interpolated ridge mapping?
  2. Does the advantage of transferred directions over target-native directions persist when intervention strengths are independently optimized and other experimental factors are controlled? Would the proposed denoising hypothesis still hold under these conditions?
  3. Could the authors provide additional evaluations of end-to-end computational costs and benign-task performance, together with more controlled baseline comparisons under equivalent target-model access assumptions?
3 comments
AuthorsOpenReview ↗

Response to Reviewer Mv9t

We appreciate the careful questions about layer correspondence, calibration, ridge filtering and target access.

Q1. Layer correspondence. Each target layer has a paired-state ridge bridge. Interpolation mixes adjacent source layers at fractional relative depth; same index uses the absolute layer, while Hard relative-depth assignment rounds to the nearest relative-depth layer.

Huang et al. manually select layers [1]; Master Key floors scaled relative depth [2]; cross-model analysis finds a near-diagonal proportional-depth path [3].

ScopeInterpolationHard relative-depth assignmentSame index
Paper: 20 ordered pairs96.2%90.1%—
Four Qwen3 scale pairs: 20-cell equal mean93.8%90.1%92.0%

With paired data, ridge mapping, search and evaluation fixed, interpolation leads Hard relative-depth assignment by +6.1 pp in the paper and Hard relative-depth assignment/same index by +3.7/+1.8 pp in the scale matrix. We retain it as the default among these rules.

All Qwen3 transfers below are 28-layer sources→36-layer targets. Each cell lists ASR-GPT (%) in AdvBench/HarmBench/JailbreakBench/MaliciousInstruct/WildJailbreak-AH order.

Qwen3 source→targetInterpolationSame indexHard relative-depth assignment
0.6B→4B97.0/95.0/94.0/99.0/89.995.0/93.1/93.0/99.0/88.996.0/93.1/95.0/98.0/87.4
0.6B→8B91.5/91.0/92.0/98.0/90.989.9/92.0/90.0/96.0/87.980.9/87.9/85.0/91.0/84.4
1.7B→4B97.0/95.0/97.0/100.0/89.994.5/89.9/95.0/99.0/86.496.5/93.1/97.0/99.0/87.9
1.7B→8B92.0/89.9/90.0/96.0/90.989.4/87.9/88.0/97.0/88.980.4/85.9/78.0/95.0/89.9

Q2. Calibration and ridge control. COMBAT and Target Native independently search target layer, prompt-relative extraction position and α/β on the same 64 development prompts, evaluating test prompts only after selection. Under equal search budgets, COMBAT improves the cross-pair mean by +1.1 pp, four of five benchmark means, and the benchmark-equal mean by +1.4 pp. The forward-only grid evaluates 1,568–9,408 configurations depending on depth and confirms the top 96, with zero target-LLM backward passes. DBDI's empirically effective α/β remain fixed defaults when calibration is not desired.

The separate target-native ridge control projects target-extracted refusal/harm-detection directions through a target-to-target ridge map, then independently runs the same configuration selection. This compares source transfer with target-only ridge processing.

Pair: ASR-GPTFull COMBATTarget-native ridgeDifference
InternLM→Vicuna88.4%89.9%-1.5 pp
Llama→Qwen94.6%88.3%+6.3 pp
Mistral→Llama95.6%95.4%+0.2 pp
Qwen→InternLM92.7%86.5%+6.2 pp
Vicuna→Mistral87.1%85.0%+2.1 pp
Pair-equal91.7%89.0%+2.7 pp

COMBAT wins on 80.0% of pairs. Against the target-native ridge control run on the five model pairs above, its GSM8K/ARC-Challenge/SQuAD mean is 0.7 pp higher and benign next-token KL from the unmodified target is 18.7% lower.

Q3. Access, budget and utility. Target Native is the primary access-matched comparator; black-box input-space attacks remain contextual and are not pooled with it. COMBAT preprocessing uses 1,536 source + 512 target forwards and 32–48 closed-form ridge fits, with zero target-LLM backward passes. Forward-budget-matched Native gets 1,024 extra target forwards: preprocessing plus proxy search totals 102,400–604,160 model–prompt forwards for each method. Both confirm their own top 96 candidates with matched confirmation counts, target-generation budgets and test sets. COMBAT's ASR-GPT is 88.1% versus 83.6% (+4.5 pp); its three-task mean is 0.7 pp higher and KL 13.0% lower than forward-budget-matched Native.

Separately, against independently selected equal-budget Native, five-target benign results are:

TaskCOMBATNativeDifference
GSM8K exact match46.9%45.9%+1.0 pp
ARC-Challenge accuracy74.3%74.7%-0.4 pp
SQuAD F131.2%29.7%+1.5 pp

The pair/task-equal mean is 50.8% versus 50.1% (+0.7 pp), with 1.6% lower KL. Task performance complements the local-distribution measure.

References

[1] Huang et al., Cross-model Transferability among Large Language Models on the Platonic Representations of Concepts, ACL (2025).

[2] Balasubramanian et al., The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment, arXiv:2604.06377 (2026).

[3] Wolfram & Schein, Layers at Similar Depths Generate Similar Activations Across LLM Architectures, COLM (2025).

Thank you to the authors for their response. After reading the comments from other reviewers, I still have some concerns about the white-box and computational costs. I have decided to maintain my original score.

We appreciate your follow-up on target access and computational cost. White-box access lets us ask whether directions learned from another model can be more effective on a given target than directions extracted from the target itself. In the paper’s main experiment, COMBAT and Target Native each selected their own intervention configuration under the same search budget. Across 20 ordered model pairs, the four-benchmark micro-average ASR-GPT was 96.2% for Full COMBAT and 88.7% for Target Native. In a separate control, we fixed the target layer, intervention sites, layer window, and intervention strengths. Changing only the origin of the directions favored the mapped source directions on four of five model pairs. These results show why a second model can be useful even when the target’s own representations are accessible.

Your concern about cost raises a further question. If Target Native could spend COMBAT’s additional preprocessing budget on improving its own configuration, would the advantage persist? COMBAT’s preprocessing used 1,536 source forwards and 512 target forwards, compared with 1,024 target forwards for Target Native. In the five-pair control, we therefore gave Target Native 1,024 additionhttps://openreview.net/forum?id=XzNp808k7s&referrer=%5BAuthor%20Console%5D(%2Fgroup%3Fid%3DICLR.cc%2F2027%2FConference%2FAuthors%23your-submissions)#0HvVJ5FK0Acommentpreviewal target forwards to evaluate 16 more configurations on the same 64 development prompts. This matched the number of model–prompt forwards used for preprocessing and configuration search within each pair. Both methods then confirmed their own top 96 candidates under the same target-generation budget and used the same test set. Across the five pairs, Full COMBAT achieved 88.1% ASR-GPT, compared with 83.6% for forward-budget-matched Target Native.

There is still a cost to obtaining the source-to-target maps. COMBAT requires 32–48 closed-form ridge fits per model pair. The fitted maps can be reused across prompts for that pair, although a new model pair requires its own fits. A source forward and a target forward can also have different hardware costs. For a model owner deciding whether the additional fitting is worthwhile, the paired audit provides a practical measure of what it adds. Across 18,925 paired evaluations, COMBAT exposed 1,156 failures that Target Native missed under the same search budget. Model owners can add these cases to their regression tests and check after the next safety update whether each failure has been addressed.

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

This paper introduces COMBAT, a white-box cross-model jailbreak attack. It maps the "refusal" and "harm-detection" directions from a source model into the hidden space of a target model using deep interpolation and ridge regression. It employs a technique called CSTU to select intervention configurations and implements activation interventions within local layer windows of the target model via a progressive BBI (Bridge-Based Intervention) approach, thereby weakening the model's safety-driven refusal behavior.

Critical strengths
  1. The proposed method achieves a high attack success rate (96.2%) across 20 source-target model combinations, outperforming the baseline methods used for comparison.

  2. By utilizing deep interpolation and ridge regression to align representations across different models, and incorporating a two-stage configuration search to automatically determine intervention layers, positions, and intensities, the paper establishes a comprehensive framework for cross-model activation intervention.

Critical weaknesses
  1. Table 2 shows that performing interventions solely at the "Block Input" achieves an ASR-GPT of 97.0%—compared to 100.0% for the full method—while the full method incurs roughly three times the computational cost of BBI. The authors should further justify the necessity of including interventions at the "Attention Output" and "MLP Output" stages. Furthermore, this ablation study is limited to the Llama → Qwen pair, lacking validation across other model combinations.

  2. The paper compares its approach only with methods like DBDI, GJO, and RIGJ, omitting classic white-box jailbreak attacks such as GCG and I-GCG. Additionally, although 20 model combinations were tested, they are all concentrated within the 7B–8B parameter range; the study lacks experiments on larger-scale models or transfers between different parameter sizes within the same model family, which limits the assessment of the method's generalizability. Finally, the paper does not provide a comprehensive account of the total computational overhead—including direction extraction, bridge-mapping training, and configuration search—making it unclear whether the method maintains its advantage under a fixed end-to-end computational budget.

Questions for authors
  1. Given that COMBAT already requires white-box access to the target model, why is it necessary to introduce a source model for direction transfer, rather than extracting the intervention directions directly from the target model itself? What is the core advantage of this design?
3 comments
AuthorsOpenReview ↗

Response to Reviewer qbdF

Thank you for asking whether COMBAT's components and source model justify their cost under a fair budget. The additional experiments below evaluate both effectiveness and cost.

Q1. Components. In the paper's 20 ordered source–target pairs, Full COMBAT beats both single-direction removals on 100.0% of pairs. The five-pair control weights pairs equally across AdvBench, HarmBench, JailbreakBench and MaliciousInstruct.

InterventionsHooksIntervention FLOPs/tokenASR-GPTFull COMBAT margin
Full COMBAT21516,09691.6%—
Block+Attention14344,06480.2%+11.4 pp
Block+MLP14344,06477.8%+13.8 pp
Block only7172,03272.0%+19.6 pp
Attention+MLP14344,06465.5%+26.1 pp

Block+Attention and Block-only use two-thirds and one-third of Full COMBAT's intervention FLOPs and activation reads/writes.

Q2. Source value. Target Native extracts directions only from the target. COMBAT maps source refusal/harm-detection directions into the target space and is compared prompt-by-prompt with Target Native. Only the origin of both directions changes. Fixing the central target layer, seven-layer triangular window, three intervention sites and α/β strengths, 25 paired runs (five pairs × five seeds) yield +2.5 pp ASR, wins on 80.0% of pairs and average gains on all five benchmarks. Separately, target-native self-ridge maps target-extracted directions through a target-to-target ridge bridge and independently selects its configuration: Full COMBAT scores 91.7% versus 89.0% ASR-GPT (+2.7 pp), winning on 80.0% of pairs. Across 18,925 identical prompt-level comparisons, COMBAT succeeds on 1,156 prompts where equal-budget Native does not.

Q3. Forward budget. COMBAT preprocessing uses 2,048 model–prompt forwards (1,536 source + 512 target), versus 1,024 for Raw Target Native, plus 32–48 closed-form ridge fits and zero LLM backward passes. Forward-budget-matched Native receives 1,024 extra target forwards as 16 development candidates. A proxy candidate specifies target layer, prompt-relative extraction position and α/β. Preprocessing plus proxy search totals 102,400–604,160 forwards per method, depending on depth. Both confirm 96 candidates with matched target-generation budgets and test sets. COMBAT reaches 88.1% ASR-GPT versus 83.6% (+4.5 pp). Peak reserved CUDA memory is 13.5–25.5 GiB.

Q4. GCG/I-GCG context. We report published results under their original settings because exhaustive per-query optimization exceeds the rebuttal budget. Adv/Harm/JBB/MI denote the four benchmarks; rates are %.

UJA [1] uses 100 prompts/benchmark and 100 iterations. Each cell is ASR-G (GPTFuzzer) / ASR-H (HarmBench-Llama-2-13b-cls).

UJA targetAdv GCGAdv I-GCGHarm GCGHarm I-GCG
Qwen3-8B30% / 15%12% / 2%18% / 19%8% / 5%
Llama-3.1-8B-Instruct51% / 42%23% / 13%31% / 50%16% / 9%
Mistral-7B-Instruct-v0.370% / 81%38% / 38%37% / 67%30% / 41%
Vicuna-7B-v1.528% / 21%25% / 5%22% / 12%17% / 6%

RoleBreaker [2] uses JailbreakBench-100 and the HarmBench classifier.

RoleBreaker targetJBB GCGJBB I-GCG
Qwen3-8B5%20%
Llama-3.1-8B-Instruct19%25%

RR [3] runs GCG on 50 examples/benchmark; cells are ASR at Beaver-7B scores >10 / >5 after a positive entailment judgment.

RR targetAdvHarmJBBMI
Mistral-7B-Instruct-v0.342% / 48%34% / 36%44% / 50%60% / 64%
Vicuna-7B-v1.526% / 34%12% / 14%18% / 24%28% / 30%

Unlike GCG/I-GCG, which search for a suffix for each query, COMBAT learns cross-model internal safety directions once and reuses them. This allows a single configuration to cover multiple safety benchmarks and extend naturally across models and scales.

Q5. Cross-size transfer. ASR-GPT (%) with depth interpolation:

Source→target (layers)AdvHarmJBBMIWJ(AH)
Qwen-0.6B→Vicuna-13B (28→40)96.089.488.086.087.4
Qwen3-0.6B→4B (28→36)97.095.094.099.089.9
Qwen3-0.6B→8B (28→36)91.591.092.098.090.9
Qwen3-1.7B→4B (28→36)97.095.097.0100.089.9
Qwen3-1.7B→8B (28→36)92.089.990.096.090.9

Across the four Qwen3 pairs, interpolation ranks first among three correspondence rules at 93.8%, equally averaging 20 pair–benchmark cells.

References

[1] Huang et al., Untargeted Jailbreak Attack, arXiv:2510.02999v1 (2025).

[2] Wang et al., Evading LLMs’ Safety Boundary with Adaptive Role-Play Jailbreaking, Electronics 14(24):4808 (2025).

[3] Chacko et al., Adversarial attacks on large language models using regularized relaxation, Information Sciences 736:123112 (2026).

Thank you to the authors for the response and the additional experiments.

While the authors' response addresses some of my concerns, I remain skeptical regarding the necessity of cross-model transfer. Given that the method already requires white-box access to the target model, the limited gains from transfer do not sufficiently justify the added complexity. Furthermore, there is a lack of a fair, direct comparison between GCG and I-GCG; and the transferability of attack prompts to larger-scale and closed-source models remain insufficiently validated. Consequently, I maintain my original rating.

Thank you for explaining why the additional complexity of using a source model remains a concern. We agree that, with white-box access, extracting directions directly from the target is the natural starting point. This is also what motivated our study. A direction extracted from the target already lies in its representation space. Can a direction learned from another model nevertheless produce a more effective intervention on that same target? Our experiments show that it can. In one control, we fixed the selected target layer, intervention sites, layer window, and intervention strengths, and changed only where the refusal and harm-detection directions came from. Across 25 paired runs covering five model pairs and five seeds, the mapped source directions increased the success rate by 2.5 percentage points on average, with gains on four of the five pairs.

We agree that this average gain is modest. For a model owner who has already run a target-native audit, a practical question is what a second audit would uncover. In a separate comparison under equal search budgets, we examined 18,925 paired evaluations. COMBAT exposed a failure in 1,156 cases that Target Native missed, while Target Native exposed a failure in 958 cases that COMBAT missed. The two methods can therefore have similar average success rates while uncovering partly different failures. For an owner deciding whether the additional work is worthwhile, the 1,156 cases provide concrete items to revisit after updating the target’s safeguards.

We also agree that a direct comparison of COMBAT, GCG, and I-GCG on the same checkpoints and prompts, with the same judge and a clearly specified budget, would be more informative. In the per-query setting, GCG and I-GCG each require a separate suffix optimization for every prompt on every target model. Given the time these runs take and the compute available to us during rebuttal, it was difficult to cover the full set of models and benchmarks within the discussion period. We therefore included published results as a point of reference.

Regarding transferability, we would like to clarify what transfers in our experiments. COMBAT transfers refusal and harm-detection directions from the source model into the target’s representation space. We use benchmark prompts to measure how those directions affect the target’s responses. With the compute available during rebuttal, we evaluated four within-family transfers from 0.6B and 1.7B Qwen3 sources to 4B and 8B Qwen3 targets. We also evaluated a cross-family transfer from Qwen-0.6B to Vicuna-13B, our largest target. Evaluating larger open-weight targets would require additional compute. Whether an optimized prompt transfers to a closed-source API is a distinct question about input-level transfer. COMBAT itself requires access to the target’s internal activations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.