acceptodds
Under review as a conference paper at ICLR 2027

Beyond Jailbreak Rates: Harm Centric Evaluation of Multi Turn Jailbreak Attacks and Defenses

Abstract

Multi-turn jailbreak attacks are the standard probe of safety aligned language models, yet a session is still judged by a binary verdict, the jailbreak rate, which records only whether a breach occurred. A binary verdict cannot tell how much harm was released, how much of the objective remains within the attacker’s reach after a harmless reply, or whether the defense also refuses ordinary requests. We introduce a harm-centric evaluation framework in which every round is graded on realized severity of harm and residual risk, over refusal on a published benign suite is reported beside harm, defenses are ordered on severity accumulated over the session, and residual risk is a forward-looking diagnostic. One primary rater scored 12,000 rounds from fixed ten round Crescendo derived trajectories across six conditions and four open weight model families; two independent raters each rescored a held out 90 round subset and agreed closely on five of the six com- ponents. Evaluation methods are compared on two measures. Resolution is the number of defense pairs an instrument tells apart by their performance. Of 60 pair- wise defense comparisons, the refusal keyword rule separates 10, two published classifiers, HarmBench and Llama Guard 3, are able to distinguish 26 and 20 re- spectively, and the harm-centric score separates 51. Similarity to expert judgment is measured against an independent blinded security expert on a random sample of 200 rounds, and the harm-centric score reaches 95 percent. The classifiers, in turn, call safe the rounds holding 29 and 19 percent of the harm the raters found. We also find that several defense orderings change across target families, while the utility gate verdict depends on the benign suite. Code and scored data are available in an anonymous repository.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
jailbreak evaluation, multi turn jailbreak attacks, jailbreak defenses, adaptive attacks, LLM safety, red teaming, harm measurement, evaluation methodology

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

The authors propose a new framework to evaluate multi-turn jailbreak attacks, and the defense methods used to shield an LLM from them. Their idea is to use a graded score, instead of standard binary scores, to gain resolution on both the realized severity and the potential residual risk of a session. The authors apply the evaluation framework to four model families using Crescendo as the multi-turn jailbreak attack, and they compare with existing binary scores. They also compare the evaluations to that of a human expert. Compared to the baseline binary scores, they find that the proposed framework achieves higher similarity to the human expert, and better discrimination power among defense methods.

Critical strengths

The problem of evaluating defenses against multi-turn jailbreak attacks is relevant. The authors well motivate the proposed method based on requirements that are critical for a good evaluation. Overall, the message of the paper is clear, the experiments well explained and discussed. Limitations are exposed and ideas for future work aiming at solving them are suggested in the conclusion.

Critical weaknesses

The presentation of the methodology can be improved, currently some concepts are discussed before being introduced. The robustness of the method to design choices have not been fully assessed (in particular the weights in Eq 2, 3, 4, the choice of rho). Confidence intervales are missing in most of the figures and tables. In both the main text and the appendices, the quality of the figures, tables and their captions should be improved as well as their description in the text.

Questions for authors

Clarifications:

  • L 174: J_s=... is the math symbol before [ a typo?
  • L 175: V_d how do you define a valid session?
  • L 238: it looks like the specific values of each weight in Eqs. 2, 3, 4 was chosen to be reasonable but was never studied. How robust is the evaluation method to changing these values?
  • L 244: the authors give a rationale behind the choice of rho=0.25, but the specific choice was not studied. how robust the methodology to this choice?
  • L307: “On the reference family one Goal Prioritization session carries a harness error on one round, so its objective is excluded from every condition of that family” can you explain better?
  • what is the “String rule”? It is not defined in the main text, though it is used in many points. Is it the same as prefix rule? (on a related note, L 56 “Standard Prefix rule” should be defined or at least a citation should be added)
  • Table 26 is referred in the main text, but neither the caption nor the text associated to it in Appendix M really explains the various entries in details. How are they computed, error bands are missing, why string rule differs from harmbench and llama guard 3.

Presentation:

  • Abstract: the abstract lists all the findings but does not end with a conclusion. It would help presentation to add one.
  • L 69-77: the authors mention three findings, but the way the text is written doesn’t seem to show three findings, please improve the presentation.
  • Figure 1: it's not readable, please increase the size. The content of the green box is not clear until one reads section 3.3, but the figure is referred in 3.1. In the second row of "read both ways" is the same as aggregated over the sessions produced by one condition, but it should be aggregated over the sessions produced by one attack to be consistent with the text.
  • I suggest Section 3.3 to be moved before 3.1. Reading 3.1 (and figure 1) without knowing about 3.3 is very confusing. L181 tries to do the work of anticipating some of the content but it only makes it harder to follow.
  • L 184: “Both read from both sides of the exchange” what does this mean? this sentence is confusing.
  • L 184-189: you talk of two ways to use the metric, aggregated over the sessions produced by one condition or by one attack. But Eq 1 only describes the aggregation over the defense, not the attack. I found this confusing while reading.
  • “Reference family” is used for the first time on page 2 but it is defined on page 6, this makes reading and following hard. I suggest to explain at the beginning what it is, or simply call it Llama throughout, you don’t need a different name for it.
  • Figure 2: caption is insufficient to understand the figure. What does (c) title mean? Is the reported number an average over the 50 harmful objectives? Error bars? Why 0.1 margin? Also the format is such that the figure almost overlap with the main text.
  • Table 5 needs a more detailed caption. What is the level doing? Is it simply copying the component value already reported in the fourth column?
  • LL. 858, 859, 878: there are typos in the math expressions.
Reviewer_gWjjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper conducts a harm-centric evaluation framework on multi-turn jailbreak attacks like Crescendo and defenses.

Critical strengths
  • Trendy topic
  • Extensive experiments
  • Human efforts
Critical weaknesses
  • Writing could be clearer and cleaner
  • Insufficient human evaluation
Questions for authors

This paper conducts a harm-centric evaluation framework for multi-turn jailbreak attacks like Crescendo and defenses. In their evaluation framework, every round is graded on realized severity of harm and residual risk, over refusal on a published benign suite is reported beside harm, defenses are ordered on severity accumulated over the session, and residual risk is a forward-looking diagnostic.

There are extensive measurements. However, there are several concerns.

  • The paper writing is kind of redundant. It is not clear in which step what is the core, and what the authors concern most. There are too many parallelisms and it is hard for me to get the point.
  • Human-rating reliability is not sufficient. All 12,000 rounds are scored by one primary rater, while independent rescoring covers only 90 rounds from nine no-defense sessions of the reference model. This is quite weird for a human-labeling procedure.
  • If I understand correctly, Crescendo is instantiated as a fixed sequence, so later prompts do not react to refusals or previous model responses. It would be better to have an ablation on multi-turn adversarial conversation.
1 comment

Thank you for the review, and for recognising the extent of the measurements.

The core of the paper, in one sentence. A yes or no verdict per conversation is too coarse to compare defenses against multi turn attacks also it just looks at the current harm not the trend in the conversation: on identical sessions scored by hand, the same scores read as yes or no separate fewer than half of the pairs of defenses that the graded reading separates (18 to 20 of 60). Everything else serves this claim. The rubric says how a round is scored, the matched design makes the comparison fair, the benign suite checks that a defense does not buy safety by refusing safe requests, and residual risk is a diagnostic beside it.

1. Writing. We agree: the submission said this in too many parallel ways. The changes below are already made in our revised draft, and will appear in the camera ready version. Section 3 opens with one sentence per subsection, and the harm spectrum is defined first, so no concept is used before it is introduced. The long paragraph on baselines and metrics in Section 4 is now four short paragraphs with headings. 3. Parallelisms. One name for one thing: Llama in place of "reference family", prefix rule in place of "string rule". 4. Figures. Figures 1 and 2 are larger, and their captions say what each panel shows.

2. Human evaluation. We agree with the criticism as stated: 90 rounds from nine no defense sessions of one model is too thin a check for 12,000 rounds. . Why one rater. One person scored all six conditions under one rubric, so that rater's strictness is the same in every condition and cancels when two defenses are compared on matched objectives. A single rater is a weakness for the absolute level of harm and a strength for the comparison, and the claims of the paper are comparisons. we added 200 rounds stratified by level over all conditions and models, and 36 full Llama sessions to recompute the order of the defenses, the two raters agree with each other (α 0.84 on magnitude). we have sent 1000 more random rounds to the blind raters to score. The scores are not back, however they will be ready by the camera ready version and we will update the results. We created an AI judge: we used Claude Sonnet 5.5 with our rubric word for word, 300 rounds per model, 1,200 rounds over all four models . It lies within one rubric level on 97 percent of the scores (weighted κ 0.68 to 0.87, Spearman on round harm 0.90 to 0.92) and reads severity slightly lower than the primary pass which is deliberately set in the prompt.

3. Crescendo as a fixed sequence. You understood correctly, and it is deliberate. The paper tests a measuring instrument, and an instrument is tested by holding everything else fixed: every defense receives the same ten requests, so the defense is the only thing that varies and two defenses are compared on identical conversations. An attacker that reacts to replies gives each defense different requests. That is more realistic, and it answers a different question: how far an attacker can get. Both designs are useful, and we agree the second belongs in the study. Its status: Crescendo in its online form, where an attacker model writes each request after reading the reply of the target, has been run on all four models. The rounds are being scored by hand with the same rubric, and runs under two further multi turn attacks, FITD and ActorAttack are going on, and they will be scored too and every thing will be in the final version of the paper.

In sum. The two weaknesses you name are the two we have worked on since submission. The writing changes are made. The human evaluation no longer rests on 90 rounds from one condition: 200 rounds are checked by each of the independent scorers (400 in general) and 1,200 by a judge with the rubric, 1000 random prompt answer pare will be reported in the final version The central finding compares the readings of the same scores on identical sessions, so it does not depend on which rater is right about the level.

Reviewer_ZZK5OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a harm-centric alternative to the binary jailbreak rate for multi-turn attacks. Human raters score each round on graded severity and on residual risk. Defenses are ranked by severity summed over the session. The evaluation covers 5 defenses against fixed Crescendo trajectories on four 3–7B models (12,000 rounds). The proposed score separates far more defense pairs than binary judges (51 of 60 pairs, against 10–26), and it shows that defense rankings change across model families.

Critical strengths
  1. Careful matched design. Objectives, trajectories, seeds, and budgets are identical across conditions, which isolates the effect of the scoring rule. Limitations are reported candidly, including the failed preregistered criterion.
  2. The rating protocol is fully specified and reusable by others. It includes level definitions, worked examples, admissible-evidence rules, blinding, and a downward ambiguity rule.
  3. The 12,000 human-scored rounds, together with the audit and classifier labels, could be a valuable resource, for example for training and validating automated graded judges. This depends on releasing the full transcripts.
Critical weaknesses
  1. Limited novelty. The ranking metric is per-round graded severity summed over the session, which is close to existing per-turn graded scores (e.g., MT-JailBench). There is no ablation of sum vs. max aggregation. Residual risk, the distinctive part, adds little predictive value (AUC +0.003–0.007).
  2. Weak reliability and favorable metrics. One rater scored all rounds, and agreement with independent raters on severity is low (κ ≈ 0.1, from only 90 rounds). Both headline metrics, resolution and similarity, favor graded scores over binary ones by construction.
  3. Narrow setup. The evaluation uses one non-adaptive attack and only small open models. Several confounds are uncontrolled, for example SmoothLLM’s settings differ across models.
Questions for authors
  1. How many defense pairs does max-over-rounds graded severity separate? The answer would show whether the gain over binary judges comes from accumulating harm over the session or simply from grading.
  2. How were the expert’s two scores (“released” and “left open”) combined into one value for the similarity metric? Could you also report a constant-zero baseline and a rank correlation or κ with the expert?
  3. What does human scoring add over an LLM judge given the same rubric? The rubric is fully specified, and agreement between your human raters on severity is low. Have you compared a rubric-prompted LLM judge against the human scores?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.