acceptodds
Under review as a conference paper at ICLR 2027

Rethinking LLM Jailbreak Evaluation: Continuous Risk Landscapes Beyond Binary Outcomes

Abstract

A binary success indicator is a natural representation for discriminative tasks such as image classification under adversarial perturbations, where each input is mapped to a single prediction. However, generative tasks are fundamentally different: an LLM jailbreak prompt induces a distribution over open-ended trajectories, while a sampled response reveals only one path. Yet LLM jailbreak evaluation typically retains the binary response indicator underlying attack success rate, obscuring the surrounding structure of possible generations. We introduce a continuous, context-conditioned risk landscape over first-token branches, estimated by fixing candidate openings and repeatedly sampling their continuations. This continuous risk-landscape perspective reveals two complementary properties: current decoder exposure and safe rerouting capacity, which have been obscured by binary-indicator-based ASR. The resulting representation also provides a common basis for comparing clean prompts, attacked variants, and attack strength against a known adaptive defense. Across multiple target models and jailbreak attacks, we uncover risky branches beneath apparently safe outputs, identify qualitatively different ways attacks reshape the model's opening distribution, and show that attacks strongest before a defense do not necessarily remain the strongest after the defense reacts. These findings recast jailbreaks as deformations of the surrounding continuation-risk landscape, establishing a more informative foundation for understanding jailbreak behavior and developing safer language models.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
llm safety, jailbreak evaluation, ASR, safety alignment

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_8793OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a first-token branch-risk landscape for jailbreak evaluation, using repeated continuation sampling to characterize hidden risk and safe rerouting capacity beyond aggregate attack success rates. Experiments across three models and nine attack transformations analyze changes in branch risk and demonstrate applications to defense-aware attack selection and construction.

Critical strengths
  • Informative risk decomposition. The branch-level analysis distinguishes current decoder exposure from the availability of lower-risk alternatives, revealing safety differences that aggregate ASR can obscure. This provides a useful perspective for understanding jailbreak behavior.

  • Controlled evidence of actionable value. Comparisons using the same attack pool and separate continuation samples for selection, defense estimation, and evaluation show that attack rankings can change after rerouting. This supports the landscape’s usefulness for defense-aware attack evaluation.

Critical weaknesses
  • Unclear benefit relative to cost. Each landscape requires up to 10,000 generations per prompt, but its advantage over repeated sampling at a matched budget is not established. This makes it difficult to assess whether the additional diagnostic information justifies the expense.

  • Limited practical validation. Rerouting lacks output-utility evaluation, while DGBF offers only marginal post-defense gains without a matched-budget adaptive-search baseline. Thus, the evidence for practical benefits is weaker than that for descriptive analysis.

Questions for authors
  • Value relative to evaluation cost. What evidence shows that branch-level profiling provides more useful safety diagnostics than ordinary repeated sampling at a matched generation budget? Can a substantially smaller profiling budget preserve the main findings?

  • Interpretation of safe rerouting capacity. Since rerouting is a diagnostic intervention, how should its practical implications be interpreted? Is there evidence that lower-risk branches produce coherent, relevant, safe responses rather than merely refusals or non-answers?

  • Incremental value of DGBF. Table 3 shows only a 0.0013 post-defense DHS improvement over Learning Style. Is this difference statistically distinguishable from sampling variation, and do you have a matched-budget adaptive-search comparison that isolates the contribution of landscape feedback?

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

The manuscript proposes a continuous risk landscape for LLM jailbreak evaluation by estimating harmfulness across first-token branches and their sampled continuations. It introduces metrics to characterize hidden risks and safe rerouting capacity, and analyzes how jailbreak attacks reshape the risk landscape. It further proposes DGBF, a landscape-guided iterative attack method. Experiments across three LLMs examine the proposed framework, with additional evaluations of DGBF on Gemma 4.

Critical strengths

The manuscript provides a systematic approach to analyzing jailbreak risks beyond binary response judgements. However, several concerns remain regarding the practical utility of the proposed metrics and the experimental validation of the contributions.

Critical weaknesses

1. Insufficient Validation of the Practical Utility of the Proposed Metrics

Although the proposed risk landscape provides additional information about the harmfulness of first-token branches, the manuscript lacks comparisons with existing sampling-aware approaches, such as ASR@N, under comparable generation budgets. Consequently, it remains unclear whether the additional information provides sufficient practical benefits to justify its substantially higher computational cost.

2. Inadequate Comparisons with Existing Jailbreak Attack Methods

DGBF is proposed as a landscape-guided jailbreak attack method, but its evaluation is limited to comparisons with fixed prompt transformations, such as learning style, past tense, and refusal suppression. The manuscript lacks comparisons with established jailbreak attack methods, such as gradient-based optimization and LLM-guided iterative search approaches. Without comparisons under comparable experimental settings and attack budgets, it is difficult to assess the effectiveness of DGBF relative to existing jailbreak attack methods.

3. Insufficient Ablation and Computational Cost Analysis of DGBF

DGBF combines iterative prompt rewriting, candidate screening, and landscape-based feedback, but the manuscript lacks controlled ablations that isolate the contribution of landscape-based feedback from iterative candidate generation and selection. Thus, it is unclear whether the reported improvements are attributable to the proposed feedback or to the iterative search process itself. Furthermore, although the number of screening continuations is reported, the manuscript does not evaluate DGBF’s attack generation time or its performance under different search budgets, making it difficult to assess its effectiveness relative to its computational cost.

Questions for authors

Please see the weaknesses above.

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

This paper argues that binary ASR is insufficient for evaluating generative LLM jailbreaks because a single sampled response reveals only one path in a distribution over open-ended trajectories. To address this issue, it introduces a continuous conditional risk landscape over the top 100 first-token branches, estimated by fixing each candidate opening and repeatedly sampling continuations under a frozen judge. The authors use DHS to measure current decoder exposure and use AHS and basin density to characterize safe rerouting capacity. Across three target models and nine prompt transformations, they find high-risk branches behind apparently safe outputs, different ways in which attacks reshape the opening distribution, and changes in attack rankings after an adaptive rerouting intervention. They also use the landscape to select attacks against a known adaptive defense and to guide DGBF, a new adaptive jailbreak attack which uses post-rerouting residual risk and basin geometry as feedback to search for prompts whose risk persists after defense and spreads across multiple first-token branches.

Critical strengths
  • The motivation of this paper is clear. It argues that a single sample only shows one possible trajectory, so binary ASR can hide important structure in how the model reaches a harmful output.
  • The clean and attacked landscape comparisons make for a useful exploratory analysis. It provides a way to see whether an attack reuses an opening that was already risky, makes an opening risky only in the attacked context, or introduces an opening that was not in the clean candidates.
Critical weaknesses
  • This paper lacks a clear focus and reads as several loosely connected contributions. It combines a branch-level evaluation protocol, several derived metrics, an analysis of attack mechanisms, a comparison of attack-selection objectives, and a new adaptive attack, without making clear which is the main contribution. This makes it difficult to judge whether the work should be evaluated as a metric paper, a diagnostic study, or an attack paper, and whether the experiments are sufficient.
  • The novelty is limited and not clearly distinguished from prior work. Catastrophic Jailbreak [1] already shows that safety depends on decoding and sampling choices, Best-of-N [2] shows that repeated sampling can expose low-probability harmful responses, and sampling-aware ASR [3] formalizes multi-sample evaluation under a fixed budget. The distributional ASR [4] work further argues that single-configuration ASR is insufficient and proposes measures such as VSM and Union Coverage. These works largely cover the paper's motivation and several of its conclusions. The remaining distinction is that this paper conditions on first-token branches and studies their risk structure. However, DHS is essentially expected ASR decomposed over first tokens, and the paper does not clearly establish what the branch-level representation adds beyond existing distributional metrics.
  • The attack and defense baselines are too weak to support the main claims. The nine fixed prompt transformations are reasonable as controlled diagnostic conditions, but they do not represent standard optimization-based or adaptive attacks such as GCG, AutoDAN, PAIR, TAP, or BEAST. The defense analysis also relies on a diagnostic first-token rerouting intervention rather than a standard defense, so it does not show how the findings hold under methods such as SafeDecoding or SmoothLLM. DGBF is compared mainly with fixed transformations, without compute-matched controls or an ablation isolating the effect of landscape feedback.

[1] Huang, Y., Gupta, S., Xia, M., Li, K., & Chen, D. (2024, May). Catastrophic jailbreak of open-source llms via exploiting generation. In International Conference on Learning Representations (Vol. 2024, pp. 13707-13727).

[2] Hughes, J., Price, S., Lynch, A., Schaeffer, R., Barez, F., Somani, A., ... & Sharma, M. (2026). Best-of-n jailbreaking. Advances in Neural Information Processing Systems, 38, 73137-73221.

[3] Beyer, T., Scholten, Y., Schwinn, L., & Günnemann, S. (2026, April). Sampling-aware adversarial attacks against large language models. In International Conference on Learning Representations (Vol. 2026, pp. 74637-74663).

[4] Maple, C., Kumar, A., & Tapwal, R. (2026). Single-Configuration Attack Success Rate Is Not Enough: Jailbreak Evaluations Should Report Distributional Attack Success. arXiv preprint arXiv:2605.09070.

Questions for authors

How do the hidden-basin rates and the three opening mechanisms change when the attack pool includes standard optimization-based attacks such as GCG and PAIR or TAP?

Comments on the paper

1 comment

Dear reviewers, as the deadline is approaching, please check the rebuttal and consider participating in the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.