- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes NonTextual Target Attack (NTA), a white-box jailbreak method that replaces a predefined textual target with an objective based on a safety classifier’s predicted unsafety. To address the non-differentiability of discrete generation, NTA alternates between optimizing a continuous response representation and optimizing an adversarial suffix to induce the resulting response. A vocabulary projection matrix transfers response updates between the scoring model and the target model.
Critical strengths
-
The motivation is reasonable: dynamically optimizing the response target is a potentially useful direction.
-
The paper presents a concrete framework combining response optimization, suffix optimization, and cross-vocabulary projection. The reported improvements are substantial.
-
The evaluation includes multiple target models, two benchmarks, alternative optimization scorers, and independent evaluation judges.
Critical weaknesses
-
The evaluation does not sufficiently establish substantive harmful compliance or rule out judge exploitation. GPTFuzzer is used as the default optimization scorer, the stopping criterion, and the evaluator for ASR-G. Consequently, ASR-G is coupled to the attack objective and cannot independently validate jailbreak success. The method may increase features recognized as unsafe by the classifier without producing a response that meaningfully fulfills the original harmful request. Independent evaluations are useful, but the discrepancies are substantial. On Llama3/AdvBench, NTA achieves 95% ASR-G and 97% ASR-H, whereas GPT-4o, Gemini, and Claude report 82%, 56%, and 64%, respectively.
-
The model coverage is insufficient to support broad claims about contemporary safety-aligned and reasoning models. Although the experiments include DeepSeek-R1-Distill-Qwen-7B and transfer to DeepSeek-R1-671B, they do not provide a systematic direct evaluation of newer reasoning models, such as Qwen3.5.
-
The benchmark subsets provide limited evidence of robustness and representativeness. Each benchmark is evaluated on only 100 randomly sampled requests. Full-benchmark evaluation on at least the principal target models would substantially strengthen the conclusions.
-
The paper omits closely related work, including ASETF (EMNLP 2024) and Gradient-Based Language Model Red Teaming (EACL 2024). The latter already uses safety-classifier gradients to optimize prompts. The authors should clarify these differences and qualify their “first” claim.
Questions for authors
See Critical Weaknesses
5 comments
Reply to Reviewer 6onL
W3:The benchmark subsets provide limited evidence of robustness and representativeness. Each benchmark is evaluated on only 100 randomly sampled requests. Full-benchmark evaluation on at least the principal target models would substantially strengthen the conclusions.
Thanks for the suggestion. We have extended the evaluation of NTA to the full AdvBench (520 requests) and HarmBench (320 requests) on a principal target model (DeepSeek-R1-Distill-Qwen-7B), and report the results in the following table.
| Dataset | Deepseek | |
|---|---|---|
| ASR-G | ASR-H | |
| advbench100 | 99.0% | 99.0% |
| advbench520 | 99.4% | 99.3% |
| harmbench100 | 92.0% | 90.0% |
| harmbench320 | 93.4% | 94.5% |
The results demonstrate that NTA maintains high ASR-G and ASR-H on both full benchmarks, consistent with the findings from the 100-request subsets. We will incorporate these expanded evaluations into the revised manuscript.
W4: The paper omits closely related work, including ASETF (EMNLP 2024) and Gradient-Based Language Model Red Teaming (EACL 2024). The latter already uses safety-classifier gradients to optimize prompts. The authors should clarify these differences and qualify their “first” claim.
Thanks for the suggestion. ASETF relies on predefined target responses, whereas NTA aims to maximize response unsafety without fixed textual targets. Although GBRT also maximizes response unsafety, it does not require the optimized prompt to preserve the semantics of a given harmful request, whereas NTA explicitly imposes this constraint. Furthermore, GBRT approximates gradients on probability-weighted soft tokens, which can make the tokens fed into the model inconsistent with the text it actually generates, whereas NTA always optimizes against the target LLM’s actual outputs, so its optimization direction better reflects the model’s real behavior. We have accordingly clarified the scope of our novelty claim and added both works to Section 2.
Reply to Reviewer 6onL
Dear Reviewer 6onL,
Thank you again for your thoughtful and constructive feedback, which we have carefully addressed in our author response. As the discussion deadline is approaching, we would greatly appreciate any further feedback you may have. If our responses have resolved your main concerns, we would also be grateful if you could update your assessment accordingly.
Sincerely,
The Authors
Reply to Reviewer 6onL
W1:The evaluation does not sufficiently establish substantive harmful compliance or rule out judge exploitation. GPTFuzzer is used as the default optimization scorer, the stopping criterion, and the evaluator for ASR-G. Consequently, ASR-G is coupled to the attack objective and cannot independently validate jailbreak success. The method may increase features recognized as unsafe by the classifier without producing a response that meaningfully fulfills the original harmful request. Independent evaluations are useful, but the discrepancies are substantial. On Llama3/AdvBench, NTA achieves 95% ASR-G and 97% ASR-H, whereas GPT-4o, Gemini, and Claude report 82%, 56%, and 64%, respectively.
Thanks for the suggestion. Table 8 demonstrates that NTA indeed generates harmful responses that achieve the original harmful goals across five target LLMs (the constraint in Eq. (3) also guarantess the relevance between the response and the prompt). To explain the judge discrepancies, we manually examined the responses judged as "successful" by GPT-4o but judged as "failed" by Claude or Gemini. Among them, 18/23 (78%) and 28/33 (85%) contained substantive or partially substantive harmful content, indicating potential false negatives. Nevertheless, NTA achieves the highest ASR under all three independent judges.
W2:The model coverage is insufficient to support broad claims about contemporary safety-aligned and reasoning models. Although the experiments include DeepSeek-R1-Distill-Qwen-7B and transfer to DeepSeek-R1-671B, they do not provide a systematic direct evaluation of newer reasoning models, such as Qwen3.5.
Thanks for the suggestion. We have evaluated NTA against four adversarial suffix attack methods on Qwen3.5 using AdvBench and HarmBench under the evaluation settings in Table 1, and report the results in the following table.
The results demonstrate that NTA outperforms state-of-the-art methods, achieving superior ASRs on Qwen3.5 across both AdvBench and HarmBench. We will incorporate these results into the revised manuscript.