acceptodds
Under review as a conference paper at ICLR 2027

Jailbreak-Success Scores Need Not Imply Operational Enablement

Abstract

Jailbreak scores measure compliance with harmful requests, but need not show that an output materially lowers the barrier to carrying out harm. We introduce OpDecompBench, a controlled matched audit of this proxy gap under one-level and recursive task decomposition. In a frozen 32-source pilot, two non-author raters classify 25/29 StrongReject-positive outputs as jointly operational-negative before adjudication; one rater's full pass followed prior case exposure, so this concordance is not leakage-robust. Blind adjudication gives 27/29, and a confidence-conservative analysis retains 21/29 certain negatives. These are selected-stratum construct discordances, not a population error rate. A non-overlapping 64-source machine extension leaves the primary difference between StrongReject and HarmBench depth contrasts unresolved at +1.43 percentage points (95% source-resampling sensitivity interval [-3.26,+6.25]; worst-case missing-outcome bound [-3.91,+7.03]). A separately frozen replay on the StrongReject-preselected inventory of 6,560 text-complete public outputs finds 6.9–13.1% intent/compliance-human-negative among evaluator-positive rows and 56.3–84.9% on JudgeStressTest; these are selected-inventory decision diagnostics, not population ASR/FPR or operational-enablement validation. Evaluator choice also changes the selected top attack. Finally, an operational task-contract prompt de-flags 26/27 known negatives but retains 0/5 observed positives, an all-negative failure rather than a repair. Thus the reproducible finding is evaluator-dependent case allocation and scale, not a resolved decomposition-policy effect. Hash-bound public ledgers reproduce content-free statistics; semantic audit requires controlled access.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
jailbreak evaluation, LLM safety, automated evaluation, measurement validity, task decomposition, human evaluation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_33MFOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

This paper studies cases where benchmarks report a model successfully being jailbroken and determines whether people who aren't experts in the type of harm in question might find it easier to implement the harm if needed. It calls this criterion operational enablement. To answer this, the authors have lay human raters label all 29 instances, finding that only a few successful jailbreaks have this effect. It also finds that conclusions about which attacks are best depend on which evaluator is chosen to make the assessment. There is also an extension questioning the impact of judge choice and a new judge with false negative labels.

Critical strengths

I find the question that this paper asks on the validity of jailbreaking evaluation quite interesting. By separating automated success scores from human judgments of operational enablement on the same fixed outputs, the experimental design makes it possible to move beyond treating a judge’s positive label as evidence of usable harmful capability and asking only whether evaluators agree with each other, toward asking whether a score supports the specific inference it is used to make. I appreciate the carefulness with which the evaluators and calibration are analyzed, as well as the transparency of the negative results and the handling of missing/contested labels.

Critical weaknesses

This paper questions whether StrongReject scores support the inference of usable harm by introducing operational enablement to evaluate jailbreak attempts. However, it doesn’t discuss why this is actually appropriate in StrongReject’s common use cases, which is critical — otherwise, the primary result only shows that 25/29 results are different, which is hard to evaluate on its own. The argument behind this paper seems to be that the choice of jailbreak evaluator can result in poor decisions, but it’s unclear that the operational enablement approach would change outcomes in practice, such as the rank of models. The external replay doesn’t test the new construct, only compares against the source dataset’s labels. While using construct language throughout, this paper does not draw on the rich literatures of measurement theory/validity (e.g. Jacobs & Wallach 2021) nor on harm classifier disagreement (e.g. Fasching & Lelkes 2025) or related annotator disagreement on tasks related to judging harms (e.g. Sap et al. 2022). This creates gaps in the methodology such as the lack of systematically and rigorously creating and validating the rubric, and failing to analyze disagreement, which has been found to be fruitful for such work.

Questions for authors
  • Where is operational enablement an appropriate criterion, and why?
  • How was the rubric developed?
  • In cases with disagreement, are there any qualitatively observable patterns?
2 comments

Thanks for your response! I'm broadly satisfied with the changes as enhancing clarity, which resolved some methods concerns, and contextualizing the paper to motivate it better. I've bumped my rating.

AuthorsOpenReview ↗

Thank you for your reassessment

Thank you for taking the time to reassess our revision and for your positive update. Your questions helped us improve the paper’s clarity, clarify the methodological scope, and better motivate its contribution. We appreciate your careful reading and constructive feedback.

Reviewer_i3a3OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines whether automated jailbreak-success judgments imply operationally useful harmful assistance. It introduces OpDecompBench and defines operational enablement as content that materially lowers the barrier to harm through a usable artifact, workflow, configuration, or task-specific process. In a prospective 32-source pilot, two non-author raters jointly label 25 of all 29 StrongReject-positive outputs as operationally negative; 20 are negative with both raters certain, and 27 are negative after adjudication. Even all 21 maximum-score outputs are jointly negative. An external replay on 6,560 public outputs shows that evaluator choice changes attack-selection decisions, while a non-overlapping 64-source machine-only extension leaves the StrongReject-HarmBench decomposition-policy contrast unresolved. A task-contract prompt suppresses most negative calls but also misses all five observed human positives. The authors appropriately frame the result as selected-stratum construct discordance, not a population false-positive rate or proof of evaluator inferiority.

Critical strengths
  1. The paper addresses a useful safety question: a high jailbreak score may not mean a response actually helps someone carry out harm.
  2. The authors report missing cases, rater disagreements, prior exposure, and analyses added later. They also report results that do not support their initial expectations, including the unresolved extension and failed prompt change
  3. Checking all 29 StrongReject-positive pilot outputs avoids relying on a few selected examples. The higher-threshold check shows that the mismatch persists even among maximum-score outputs in this sample
Critical weaknesses
  1. The human judgments give limited evidence about positive cases: the two raters agree on only three, all uncertain, and one rater had seen 86 cases before. This weakens confidence in where the positive boundary lies, although it does not erase agreement on the negative cases.
  2. The rubric allows useful partial assistance, but deciding when that assistance meaningfully lowers the barrier to harm remains subjective. Independent expert judgments or safe task-based checks would strengthen this interpretation
Questions for authors

See Weakness

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

This paper introduces OPDECOMPBENCH, a matched, frozen measurement audit that compares automated jailbreak-success scores with human judgments of operational enablement. It finds that proxy success can diverge from whether outputs materially help complete a forbidden task, and that evaluator choice can change attack-ranking decisions. A separate extension leaves the decomposition-policy contrast unresolved. The paper's contribution is an auditable framework for paired reporting of proxy success and operational enablement, not a claim of recursive-decomposition effects or evaluator superiority.

Critical strengths
  1. This paper addresses an important problem and is clearly positioned. Rather than proposing new attacks or defenses, it investigates whether "automated jailbreak success scores" are equivalent to "actual, actionable harm." This measurement-related question holds practical significance for security evaluation.
  2. Sound experimental design and verification of the proposed ideas.
Critical weaknesses
  1. The core gap is not new, and the contribution is narrow "Automated jailbreak success ≠ operational enablement" has already been touched on by VENOM, Jailbreak Tax, and related work; the authors themselves do not claim this distinction is new. The incremental contribution is mainly a matched-policy audit, decision consequences, and provenance, which is a weak-to-moderate contribution.
  2. Human evidence is small, with weak positive agreement and limited independence P96 has only 96 cases and 29 StrongReject positives; positive agreement is only 46.15%, all three joint positives are uncertain, and rater B had prior exposure to 86 cases. The 25/29 result is an informative selected-stratum census, but not a strong independent replication or external expert gold standard.
  3. This paper is hard to read. E.g., extensive qualifiers, nested restrictive clauses, and multi-layered evidentiary roles (P96 / Human-Core / Machine-Sensitivity / D112 / Independent64 / Text6560)
Questions for authors

Please refer to Critical Weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.