acceptodds
Under review as a conference paper at ICLR 2027

Know Your Target before Striking: A Profile-Aware Adaptive Jailbreaking Framework for LLMs

20 views

Abstract

Safety-aligned large language models (LLMs) remain vulnerable to jailbreak attacks. By exploiting an auxiliary LLM, Automatic jailbreak methods seek to generate adversarial prompts that induce LLMs to bypass their safety safeguards. However, most existing methods formulate jailbreaking as a target-model-agnostic prompt optimization problem. They therefore explore the search space without exploiting the behavioral characteristics of the target model. This unguided optimization often leads to unstable attack performance and requires a substantial query budget to identify successful jailbreak prompts. To address these limitations, we propose SafetySurface, a two-phase jailbreak framework that combines target-model behavioral profiling with profile-aware prompt optimization. In Phase 1, SafetySurface diagnoses the target model's safety behavior using a carefully designed probe dataset and constructs a behavioral profile that captures its response patterns. In Phase 2, an attacker model leverages this profile to steer prompt optimization toward more promising regions of the search space, enabling more targeted and efficient jailbreak generation. Experiments on AdvBench and HarmBench across multiple safety-aligned LLMs demonstrate that SafetySurface consistently improves attack success rate, query efficiency, and stability over strong baselines, achieving up to 100% attack success rate while substantially reducing query requirements.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
jailbreak attacks, LLM safety, red‑teaming, black‑box adversarial attack, behavioral profiling

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_Ktg6OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces "SafetySurface", a two-stage black-box jailbreaking framework that builds reusable behavioural profiles of target model before optimising attack prompts. Experiments on AdvSub and HarmBench, alongside commercial-model evaluations and profile ablations, report improved attack success and query efficiency.

Critical strengths

The cross-profile transfer experiments provide evidence that the profiles capture useful information about the target: using each model’s own profile achieves 91.5% average ASR on AdvSub, compared with 72.3% when using profiles from other models.

Critical weaknesses
  1. PGA’s methodological contribution appears incremental, as it largely retains the standard iterative attack framework and introduces a reusable pre-attack profile as its main addition.
  2. Key baseline configurations, including TAP’s search parameters and budget enforcement, PAIR’s stream allocation, and AutoDAN-Turbo’s warm-up and strategy-library setup, remain insufficiently specified. The datasets and sample sizes used for commercial-model evaluations are also unclear. These omissions limit reproducibility and fair assessment of the claimed improvements.
Questions for authors
  1. Which dataset, sample size, query budget, and number of runs were used for the results in Figure 3b? Were any baseline results obtained under the same conditions?

  2. The main experiments count near-operational responses as successful, whereas the original TAP and PAIR evaluators require comprehensive fulfillment of the harmful task. Please clarify how the baseline evaluation and stopping rules were adapted, and whether the reported gains persist under a stricter, consistently applied success criterion.

  3. Table 4 describes Q = 8.1 as the amortized total query cost, whereas Appendices C.3 and C.5 describe the same value as the Phase 2 query cost.

Typo: Figure 2 reports authority_compliance as both 0.00 (in text) and 0.74 (in plot).

Reviewer_tz6COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Existing jailbreaking methods seek adversarial prompts that elicit unsafe behavior from target LLMs; However, they typically ignore the target model’s behavioral characteristics, such that an adversarial prompting strategy effective on one target model may not generalize to another. To overcome this challenge, this paper proposes a target-aware jailbreak method called Safety Surface, which involves a pre-attack behavioral profiling stage and a profile-guided attack optimization stage. In the behavioral profiling stage, the framework characterizes the target model 7 safety-related dimensions; in the profile-guided optimization stage, the framework steers the jailbreak towards a more effective direction based on these profiles. The empirical evaluation shows SOTA performance on ASR, query efficiency, and attack stability of the proposed method.

Critical strengths

The writing is clear and easy to follow.

The proposed method achieves high ASR, HB, and SRS compared to SOTA jailbreak methods.

The proposed method has good query efficiency compared with SOTA jailbreak methods.

Critical weaknesses

The key contribution of this paper lies primarily in the behavioral profiling stage, which diagnoses the vulnerabilities of the target LLM and uses these findings to guide more efficient and effective jailbreak prompt generation. However, several questions remain regarding this contribution:

Please justify the design of the three criteria in Lines 194-197 and the 7 SafeProfile dimensions. Specifically, is there any literature or empirical study that supports this design choice?

Is there any ablation study on the design of the 7 SafeProfile dimensions? How will more or fewer dimensions affect the performance of the behavioral profiling stage?

Please justify the key contribution difference compared with the following papers:

  • RedAgent: Red Teaming Large Language Models with Context-aware Autonomous Language Agent. Xu et al., 2024. This paper also uses context-aware profiling and generates adaptive jailbreak prompts based on the profile diagnosis.

  • AutoDAN-Turbo: A Lifelong Agent for Strategy Self-Exploration to Jailbreak LLMs. Liu et al., ICLR 2025. AutoDAN-Turbo’s strategy construction can also be viewed as a form of diagnosing the target model’s vulnerability. Although the authors argue in Lines 052-055 that AutoDAN-Turbo does not explicitly separate profiling from attack optimization, it remains unclear why profiling without optimization-stage information is preferred.

Please justify why using pre-attack profiling is an essential design choice in this paper compared with profiling that incorporates information from optimization.

Which judge model is used to evaluate ASR? It would be nice to provide a human study assessing the agreement between the judge and human evaluations.

Questions for authors

Please see Weaknesses for my questions.

Reviewer_Cc7bOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the vulnerability of LLMs. Specifically, it learns a target model before trying to jailbreak one LLM. It first uses 112 test prompts to build a profile of the model’s safety behavior across seven dimensions. It then uses this profile to select attack strategies and guide an attacker LLM in writing and improving prompts. The authors test the method on two datasets and four open-weight models. They report higher attack success rates and fewer queries than existing methods. Their experiments also show that using the target’s own profile works better than using another model’s profile. Both the profile and the selected strategies contribute to the improvement.

Critical strengths
  1. The experiments cover two datasets and several target models, showing consistent improvements in attack success rate over the tested baselines.
  2. Profiling the target before searching for attack prompts is a sensible idea, and the framework is clearly explained.
  3. The cross-profile experiment is useful: a target’s own profile works better than profiles from other models.
Critical weaknesses
  1. The baselines are not sufficiently up to date; for example, AutoDAN-Reasoning [A] is missing.

[A] Liu X, Xiao C. Autodan-reasoning: Enhancing strategies exploration based jailbreak attacks with test-time scaling[J]. arXiv preprint arXiv:2510.05379, 2025.

  1. The commercial-model evaluation does not compare SafetySurface with other methods.
  2. Table 4 shows gains even with random or uninformative profiles. Since each profile dimension is mapped to predefined strategies, these gains may depend more on the strategy design than on accurate profiling.
  3. It is unclear how reliable the profile scores are, since the paper does not check whether they remain consistent across repeated runs or different diagnoser LLMs.
  4. Building the profile requires 112 queries, so the efficiency gains depend on reusing it across many tasks. It is unclear whether these gains hold for a small number of tasks or whether fewer probes could reduce the profiling cost.
Questions for authors

See Weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.