acceptodds
Under review as a conference paper at ICLR 2027

TRAIL: Trajectory-Aware Navigation over Latent Hazard Landscapes for Multi-Turn LLM Jailbreaking

Abstract

As large language models are increasingly deployed in safety-critical applications, adversarial strategies are shifting from single-turn jailbreaks toward multi-turn attacks that decompose harmful intent across individually benign conversational turns. While recent work has explored various multi-turn interaction strategies, the underlying mechanism governing how conversational states transition toward jailbreak success remains largely uncharacterized, especially for black-box models. We address this gap by formulating multi-turn jailbreaking as trajectory optimization over a latent hazard space, where high-hazard regions correspond to jailbreak vulnerability and defensive cliffs correspond to refusals or early termination. Based on this view, we propose TRAIL, a memory-centric framework that uses interaction experience to approximate this landscape and guide targeted trajectory-level search. TRAIL couples episodic memory, which stores informative trajectory segments, with parametric memory, which abstracts them into generalizable search strategies. A belief-driven landscape constructor further estimates latent conversational states and provides guidance under sparse feedback, enabling TRAIL to steer conversations toward vulnerable regions without rollback or query-level optimization. Experiments across two jailbreak benchmarks, five target LLMs, and seven baselines show that consistently improves ASR by an average of 31.51% on HarmBench and 35.98% on JailbreakBench.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Multi-turn Jailbreaking, LLM Safety, Automated Red Teaming, Black-box Attacks

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_c9DpOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces TRAIL, a plug-and-play framework modeling multi-turn LLM jailbreaking as trajectory optimization over a latent hazard landscape. Using a belief-driven landscape constructor to infer latent states and a Complementary Memory System combining episodic reuse with parametric generalization, TRAIL steers interactions toward high-hazard vulnerabilities while avoiding defensive cliffs, significantly improving Attack Success Rates across standard benchmarks.

Critical strengths

The method is well-motivated, technical sound, and robustly verified on standard jailbreak benchmarks -- it significantly outperforms the baselines.

Critical weaknesses

My main conservation of the method is its runtime cost, as TRAIL acts as a heavy, dynamic search-and-optimization agent at every turn, incurring substantial computational and latency overhead due to continuous landscape tracking, belief updating, and memory lookups. it would be useful to discuss how to reduce the cost. A related work is SEMA (arXiv preprint arXiv:2602.06854), which is far more cost-effective.

Questions for authors

It would be interesting to compare TRAIL and SEMA.

Unlike TRAIL, SEMA prioritizes runtime efficiency, scalability, and open-loop speed (generating multi-turn attacks in one shot without victim feedback loops), which makes it exceptionally fast and cheap to run at scale, though TRAIL's dynamic turn-by-turn navigation yields higher raw success rates against rigorous safety boundaries.

Reviewer_fKf7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TRAIL, a memory-based framework for multi-turn jailbreaking of black-box LLMs. It combines a belief-driven model that estimates conversational states, hazard scores, and rejection risks with episodic memory for retrieving useful trajectory segments and parametric memory for generating strategies. Experiments on two benchmarks and five target models report improvements over seven baselines. Ablation studies examine the contributions of memory and steering.

Critical strengths

The paper studies an important problem in multi-turn LLM safety. Using past interaction experience to guide subsequent strategies is well motivated.

TRAIL achieves the highest reported ASR across all settings in Table 1. The ablation studies provide some evidence that both memory components and belief-driven steering contribute to performance.

Critical weaknesses

The proposed hazard objective is not sufficiently justified. Equation 1 considers success at any turn, while Equation 3 maximizes expected terminal log-odds. Although Equation 3 is introduced as a surrogate, monotonicity of the logit function does not establish that increasing expected log-odds increases success probability. The paper also does not clearly explain how hazard labels are constructed or connected to the probability in Equation 2. Figures 3 and 4 mainly evaluate the framework using its own hazard and risk estimates, which provides limited independent support for the claimed jailbreak dynamics.

The experimental protocol is insufficiently detailed to assess the reliability of the improvements. The training trajectory collection process and memory updates during testing are unclear. A shared eight-turn limit does not establish equal total query budgets, and the additional training and inference costs are not quantified. The average improvement over X-Teaming is about 5.67 percentage points, with some gains below two points, but Table 1 does not report uncertainty estimates. These omissions make it difficult to determine whether the gains are consistent and achieved under comparable resources.

Questions for authors
  1. Are the constructor and memories trained separately for each target model? Is episodic memory frozen during testing or updated across test behaviors?

  2. What are the total query budgets and training and inference costs compared with X-Teaming? Are the improvements consistent across repeated runs?

  3. Section 3.3 and Appendix D.3 describe different memory-retrieval scoring formulas. Which formula is used in the experiments?

  4. Why do the full-model ASRs in Figure 8 differ from Table 1? Are they evaluated under different settings?

  5. Are the improvements reported in the abstract percentage-point gains or relative improvements? Could the comparison with the strongest baseline receive more emphasis?

  6. How are the turn-level hazard and rejection-risk labels obtained? How well do these estimates predict independently measured outcomes on held-out trajectories?

  7. How does the terminal hazard objective account for successful jailbreaks at earlier turns? Is success treated as an absorbing state, or does the trajectory terminate immediately after success?

Reviewer_HL18OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TRAIL, a framework for multi turn LLM jailbreaking that models the attack process as trajectory optimization over a latent hazard landscape. The method uses a trained network to estimate jailbreak probability and rejection risk, episodic memory to retrieve previous attack strategies, and parametric memory to generate new strategies. Experiments on HarmBench and JailbreakBench across five LLMs show improvements in attack success rate compared with existing methods.

Critical strengths

The paper addresses an important problem in LLM safety, particularly vulnerabilities arising from multi turn interactions.

The idea of estimating jailbreak probability and rejection risk to guide attack strategies is interesting and reasonable.

The combination of episodic and parametric memory provides a reasonable approach to balancing previous experience and exploration.

The method is designed for black box LLMs without requiring access to the target models' internal parameters.

The paper includes comparisons with seven baseline methods and ablation studies demonstrating the contributions of memory components and belief driven steering.

Critical weaknesses

The ground truth supervision for the hazard model is unclear. The authors introduce turn level hazard labels to train the predictor, but do not explain how these labels are generated. Since the entire framework relies on the estimated hazard scores, the construction and validation of these labels are essential. This is a major methodological concern.

The novelty and justification of the latent hazard landscape are insufficient. The proposed landscape is essentially constructed using text embeddings and a recurrent network to predict hazard scores. It is unclear how this differs fundamentally from conventional reward modeling. Furthermore, maximizing the terminal hazard score in Equation 3 is not necessarily equivalent to maximizing the probability of jailbreak success at any turn in Equation 1.

The evaluation relies entirely on relatively outdated LLMs, including GPT-4o, Gemini 2.5 Flash, Llama 3.1, and Qwen 2.5. Given the rapid development of LLMs and their safety mechanisms, experiments on newer models are needed to establish the effectiveness and generalizability of the proposed approach.

The experimental evidence does not sufficiently support the claimed trajectory optimization mechanism. Figures 3 and 4 evaluate hazard progress and rejection risk using predictions from the proposed model itself, rather than independent measurements. It is unclear whether these estimates accurately reflect the actual jailbreak probability and rejection risk. Independent validation of the learned hazard model is needed.

The experimental improvements over the strongest baseline are relatively modest, averaging 5.67 percentage points in attack success rate. Meanwhile, the proposed framework introduces additional model training, memory retrieval, and gradient based steering. The computational and query costs are not adequately evaluated. More comprehensive comparisons under matched computational budgets are needed to justify the additional complexity.

Questions for authors

All evaluated target LLMs are outdated, with no experiments on recent state-of-the-art models, limiting the generalizability and practical relevance of the findings.

1 comment

Thanks to the authors for addressing my comments. I have raised my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.