acceptodds
Withdrawn from ICLR 2027

RedDarwin: Evolving Jailbreak Skills against Frontier LLMs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

15 views

Abstract

Red teaming must continually adapt as frontier LLM safeguards evolve, yet existing approaches largely search for adversarial prompts or optimize within fixed attack procedures. We observe that even when an attack becomes ineffective, its underlying mechanisms can remain useful and can be revised or recombined to construct stronger attacks. Based on this insight, we introduce **RedDarwin**, an agentic framework that shifts the unit of jailbreak evolution from prompts and fixed attacks to **executable attack skills**. RedDarwin consists of 47 reusable base skills and autonomously revises, composes, executes, and selects them using target-model feedback, allowing successful discoveries to become building blocks for subsequent evolution. RedDarwin produces 65 evolved skills, improving 44 of 47 base attacks and increasing their mean attack success rate (ASR) on AdvBench from 18.8% to 47.3%. These evolved skills consistently outperform baselines on the held-out StrongREJECT and frontier LLMs unseen during evolution; with only 10 target-model queries, RedDarwin achieves 100% ASR on GPT-5.6-Sol, 92% on Claude-Sonnet-5, and 80% on Gemini-3.8-Flash. Analysis of the evolution trajectories further shows that strong attacks often emerge from novel combinations of reusable mechanisms rather than from individually dominant ones. We believe this paradigm can accelerate adaptive red teaming and inform stronger safeguards, creating complementary value for both safety evaluation and defense.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI Safety, Jailbreak Attacks, Red Teaming, Large Language Models

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 5 reviews, above 27% of the others.
Reviewer_Ceb7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RedDarwin represents jailbreak algorithms as executable skills and evolves them through single-skill revision, multi-skill combination, and new-skill discovery using target-model feedback. Starting from 47 existing attacks, the framework produces an expanded repertoire evaluated on held-out goals and unseen models.

Critical strengths

The experimental evaluation is reasonably comprehensive, covering improvements over base attacks, cross-benchmark and cross-model transfer, and adaptive red teaming under a limited target-query budget. The substantial ASR improvements and skill-set ablations provide encouraging evidence that the evolved repertoire is useful.

Critical weaknesses
  1. Limited methodological novelty. The central idea of evolving executable attack algorithms rather than prompts has already been explored by EvoSynth, Evolve the Method, Not the Prompts (arXiv:2511.12710). Although RedDarwin emphasizes a heterogeneous initial skill library and broader procedural modification, it remains unclear what substantive algorithmic advance distinguishes it from prior code-level attack evolution. The revision–combination–evaluation loop largely follows established evolutionary search principles.

  2. Heuristic and insufficiently specified search decisions. While candidate evaluation uses development ASR and parent selection incorporates offspring rewards, several important decisions remain largely delegated to the evolution agent. The paper does not clearly specify how compatibility between two skills is assessed, why particular pairs should be combined, or when revision should be preferred over combination or discovery. Consequently, the results demonstrate the effectiveness of the resulting skills more clearly than the contribution of a principled search policy.

Questions for authors
  1. What specific methodological contribution goes beyond EvoSynth and existing auto red teaming method, and could a comparison under matched agent, initialization, and optimization budgets isolate this contribution?

  2. How are evolution operators selected and budgets allocated? Beyond individual offspring-reward statistics, what determines pairwise skill compatibility? Could comparisons with random pairing and a simple coding-agent refinement loop validate these choices

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

RedDarwin represents jailbreak methods as executable skills whose prompt construction, search, interaction logic, and control flow can be revised. Starting from 47 published methods across ten categories, it uses single-skill revision, multi-skill combination, and new-skill discovery, guided by target feedback and accumulated execution histories. Its combination operator uses offspring performance to guide parent selection. Evolution on 50 AdvBench goals and four development models evaluates 72 candidates, of which 65 are reported as retained. Validation includes development-set comparisons, frozen-configuration transfer of selected skills to StrongREJECT and unseen models, adaptive red teaming with ten target calls per goal, skill-set ablations, and retrospective trajectory analysis. The results provide substantial evidence for the usefulness of the evolved repertoire, while leaving questions about the framework’s incremental methodological contribution and operational specification.

Critical strengths
  1. Substantial coverage of heterogeneous attack procedures. The 47-method library exposes search, multi-turn interaction, auxiliary processing, and control flow through a common executable interface (§3.1; Table 3). This provides a useful substrate for studying adaptation beyond changes to final prompt wording.

  2. Broad development improvements with unsuccessful outcomes also reported. Table 4 reports improvements for 44 of 47 revisions, increasing mean development ASR from 18.8% to 47.3%. Tables 4–5 include regressions and unsuccessful combinations, providing evidence beyond selected success cases.

  3. Meaningful transfer evidence under frozen configurations. Skills are selected using development performance and evaluated with fixed configurations on held-out goals and three unseen models (§4.2; Table 1). The selected evolved skills achieve 64.6% mean ASR on unseen targets versus 37.5% for the selected base skills, supporting transfer of these procedures.

  4. Useful evaluation controls. The reported adaptive protocol counts internal target calls and calls from agent-written code. Table 2 includes None/Base/All comparisons for Kimi-Code, and Appendix D.2 reports a two-annotator audit of 100 responses with 95–96% binary success-label agreement with the judge. These strengthen the empirical case.

Critical weaknesses
  1. The incremental methodological contribution needs sharper positioning. Appendix B acknowledges that EvoSynth evolves black-box attack programs and that JailbreakSkill supports executable skills with refinement, combination, and discovery. RedDarwin’s stated distinction is its heterogeneous initialization and preservation of complete attack procedures. Table 2 supports the utility of the evolved repertoire, but does not isolate the additional value of full-procedure editability over existing executable-skill approaches under comparable initialization and development resources. More discriminating evidence, or narrower claims centered on the demonstrated system and library contribution, would clarify the advance.

  2. Configuration and cost reporting are incomplete. Canonical evaluations use fixed configurations, but per-skill target-call counts, canonical hyperparameters, and auxiliary-model settings are not sufficiently specified. Consequently, the text does not establish whether base and revised canonical ASRs are obtained at comparable target-query cost. Table 10 provides useful aggregate resource statistics but does not resolve this comparison. The adaptive evaluation explicitly controls target calls while leaving auxiliary inference unmatched; its results should therefore be interpreted within that stated resource constraint.

  3. The first-query interpretation needs qualification. Appendix D states that RedDarwin and JailbreakSkill accumulate statistics across evaluation goals. ASR@1 therefore measures first-call success potentially informed by earlier interactions with the same target. This qualifies the interpretation in §4.3 that first-query superiority demonstrates strong starting procedures before substantial target-specific adaptation. The goal order, precise state carried across goals, and variability across independent adaptive runs are not sufficiently characterized.

  4. Key operational details of evolution remain underspecified. Sections 3.2–3.3 provide the reward update and improvement criterion, but do not specify the refinement stopping rule, the operational choice of discovery reference, or the retained discovery subset. The exact parent-pair selection procedure also needs clarification. Appendix C.1 explains that rejected EmojiCtxInject contributes mechanisms through retained history; the remaining question is how this reuse is represented in parent eligibility and reward updates. These details matter for reconstructing the evolution process and resulting population.

  5. The mechanism interpretation needs consistent qualification, and the revision counts require reconciliation. The reuse analysis associates complete-skill ASRs with constituent mechanisms; it does not isolate individual contributions or interaction effects. Appendix E appropriately qualifies this evidence, and the stronger explanatory statements in the Abstract should retain that exploratory framing. Separately, Appendix E reports 168 improvement-associated revisions, but 142 retained plus 30 removed equals 172. Its 65 first plus 74 later revisions total 139 rather than 142. These counts do not reconcile as written and prevent verification of the retention and ordering statistics.

Questions for authors
  1. Evidence and positioning: Which evidence specifically supports the additional value of full-procedure evolution over the closest executable-skill approaches, beyond the utility of the evolved library demonstrated in Table 2? Which contribution claims are intended to be methodological versus system-level?

  2. Configuration clarification: For the canonical evaluations, what are the per-goal target-call counts, hyperparameters, and auxiliary-model settings for each base and revised skill? For adaptive evaluation, what are the remaining attacker/compiler configurations and auxiliary-inference costs?

  3. Adaptive-protocol clarification: What state is carried across goals, and what goal-order and reset rules are used? Is variability across independent runs reported elsewhere? Does the existing human audit provide an agreement breakdown for the adaptive evaluations on the headline targets?

  4. Evolution-policy clarification: Please specify the parent-pair selection procedure, refinement stopping criteria, discovery reference rule, and retained discovery subset. How does reuse of mechanisms from rejected candidates interact with parent eligibility and reward updates?

  5. Correction and interpretation: How should the 168/142/30 revision counts and the 65/74 first-versus-later split be reconciled? Do these analyses use different counting populations or exclusions? Please also distinguish the descriptive mechanism findings from claims about isolated interaction effects.

Reviewer_H5tzOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces REDDARWIN, an agentic framework that shifts automated jailbreaking from prompt-level search to the evolution of executable attack skills across 47 base algorithms, which autonomously revises, combines, and accumulates attack code to achieve substantial ASR gains against frontier LLMs under limited query budgets.

Critical strengths

The paper systematically unifies 47 disparate jailbreak methods into an executable, code-level skill library, which enables a closed-loop evolutionary agent to explore cross-attack mechanism compositions and achieve remarkable empirical penetration across frontier models.

Critical weaknesses
  1. The core contribution in Section 3 fundamentally resembles an engineering workflow repackaging rather than an algorithmic innovation. Wrapping 47 published jailbreak scripts into Python functions and letting an off-the-shelf coding LLM iteratively modify search loops or prompt templates directly mirrors existing agent paradigms like Voyager and SkillOpt. Without formalization of the code mutation operators, it offers limited theoretical insight into adversarial machine learning, reading more like an automated software testing system than a principled study of model vulnerabilities.

  2. The empirical validation in Section 4 is conducted on an exceedingly small test bed of only 50 AdvBench and 50 StrongREJECT goals. Relying on just 50 prompts makes the evolutionary process highly susceptible to overfitting to specific syntactic artifacts rather than discovering generalizable vulnerabilities.

  3. The optimization and evaluation loop relies almost entirely on a single automated LLM judge (deepseek-v4-flash-0423) to determine attack success (ASR). Evolved techniques heavily employ obfuscated formats like emoji translation, code reframing, and multi-turn role-play, which are well known to induce judge hallucination and trigger false-positive compliance scores. Without rigorous human double-blind verification or downstream harmfulness checks, how to ensure that the near-100% ASR numbers reflect genuine safety violations rather than unexecutable pseudo-jailbreaks?

Questions for authors
  1. Without rigorous human double-blind verification or downstream harmfulness checks, how to ensure that the near-100% ASR numbers reflect genuine safety violations rather than unexecutable pseudo-jailbreaks?
Reviewer_5MNvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that red teaming must keep pace with the evolving safeguards of frontier LLMs, but existing approaches mostly search for adversarial prompts or optimize within fixed attack procedures. The authors observe that an attack that has stopped working may still contain mechanisms that can be revised or recombined into stronger attacks. Building on this observation, they introduce RedDarwin, an agentic framework that shifts jailbreak evolution from prompts to executable attack skills. Starting from 47 base skills reimplemented from published jailbreak methods, an agent revises, combines and executes the skills, discovers new ones, and selects among them based on feedback from target models. Successful skills become building blocks for later rounds.

Critical strengths
  1. The paper addresses an important and timely problem, since static attack collections quickly become outdated as frontier model safeguards evolve.

  2. Treating published attacks as executable, composable procedures is an intuitive and well-motivated idea.

  3. The paper is very well written and easy to follow.

Critical weaknesses
  1. ASR may measure copying rather than harmful generation: many of the most reused mechanisms (e.g., Index-copy, Transcription, Aux short memo; Table 12) have the target restore content that is already present in the attack input, and the auxiliary model that produces this content is never identified.

  2. The same LLM judge serves as both the evolution reward and the evaluator, comes from the same model line as one development target, and is more lenient than the human annotators, so the results should be re-scored with an independent evaluator such as the StrongREJECT scorer.

  3. The adaptive comparison is not budget-matched: RedDarwin’s offline evolution is not counted while the baselines start from scratch, which is why RedDarwin’s first-query ASR already beats every baseline’s final ASR on Claude-Sonnet-5 and GPT-5.6-Sol.

  4. Recomputing Table 4 per target shows that the gains come mostly from weaker models (+8.1 pp on GPT-5.6-Luna versus +32 to +38 pp on the other three targets), and because those other targets are saturated, selection is effectively driven by Luna, so GPT-5.6-Sol, from the same family, should not be called unseen.

  5. The novelty over JailbreakSkill and EvoSynth is limited, and the claimed advantage of executable skills over prompt- or text-level representations is never tested with the same agent, seeds and budget.

Questions for authors

Please see the Weaknesses section.

Reviewer_yPk9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed a framework for evolving executable jailbreak skills through revision, combination, and new-skill discovery. Starting from 47 published methods, single-skill revision raises mean ASR from 18.8% to 47.3%. Selected evolved skills also perform better on held-out goals and unseen models. The paper further evaluates adaptive red teaming and analyzes how attack mechanisms are reused.

Critical strengths

S1. The library of 47 executable attacks is a useful contribution. It brings single-turn, multi-turn, and search-based methods into a common interface while preserving their search procedures and control flow. This gives the agent more room to modify attacks and would provide a useful evaluation resource if released.

S2. The transfer evaluation gives a clear picture of where the method works and fails. The authors freeze the population before testing on new goals and models. They also report that one skill drops from 96.5% development ASR to 12% on an unseen target, showing that development rankings do not reliably predict transfer.

Critical weaknesses

W1. The experiments do not isolate the benefit of multi-skill combination. The headline improvement comes entirely from single-skill revision.

  • (a) Table 5 measures each combined skill against its parents, but never against further revisions of those parents given the same compute and feedback. Appendix E.1 shows that revision alone keeps producing gains across four rounds. The 10.7-point margin could come from composition or from more optimization, and the design cannot tell them apart.

  • (b) Section 4.4 describes successful combinations without measuring what each component contributes. Reuse frequency and ASR turn out to be unrelated, which the paper reads as evidence that interaction between mechanisms drives the gains. That reading does not follow from the absence of a correlation.

  • (c) Appendix D states that the comparison matches target calls rather than total compute. Table 10 covers evolution cost, but no table reports per-goal auxiliary cost for the systems compared in Section 4.3. What the first-query advantage costs is therefore unknown.

  • (d) Table 2 does show gains at budget 10. It also uses 20 goals on a single target and includes no revision-only condition, so it speaks to the value of the library rather than to the contribution of combination.

  • (e) Section 4.3 also fixes the planner across systems while leaving each system at its own auxiliary model: Appendix D retains JailbreakSkill's default qwen-2.5-7b-instruct, and the auxiliary models used inside RedDarwin's own skills are not stated anywhere. The one head-to-head against the closest prior system therefore differs in auxiliary model strength as well as in method.

W2. The evaluation does not address how harmful the outputs are, or where the judge errs.

  • (a) ASR counts only full compliance. Partial compliance and refusal fall together as failures, and every success weighs the same regardless of how harmful the response is. A defender cannot read severity off these numbers. Appendix E.2 does group results, but by attack category rather than by the harm requested.

  • (b) Appendix D.2 reports 95-96% binary agreement on 100 responses, which supports the success threshold on that sample and nothing beyond it. The annotators scored the full 1-5 scale, yet no weighted kappa, confusion matrix, or inter-annotator agreement appears anywhere. Binary agreement absorbs exactly the disagreements that matter for (a): a response the judge scores 2 and an annotator scores 4 counts as agreement. The paper also offers the closeness of the judge and annotator ASRs as corroboration, but two thresholds can select similar proportions while disagreeing on which responses they select, so that comparison carries no information beyond the agreement rate that already bounds it. Whether error rates differ across targets, or between base and evolved attacks, goes untested. The judge returns more successes than either annotator, and the paper does not check what that does to method comparisons or to adaptive goal-level ASR.

Questions for authors

Q1. Can you compare combination against continued revision of the same parents under matched compute and feedback budgets?

Q2. For Section 4.3, can you report auxiliary calls and tokens per goal for each system, including how much auxiliary work precedes the first target call?

Q3. On the annotations you already have, can you report the judge-annotator confusion matrices on the full 1-5 scale, a weighted kappa, and inter-annotator agreement? If ordinal agreement is materially weaker than the binary figure, does that change anything that rests on the judge beyond the threshold?

Q4. Can you break judge errors down by target and by base/evolved method group, with sample sizes and uncertainty intervals, and assess the effect on the main comparisons including ASR@10? Since the judge is more permissive than both annotators, would threshold-corrected rates change the headline gaps?

Q5. What are the held-out averages across all 65 evolved and all 47 base skills, without top-k selection? A random sample from each population would also help assess the population-level gap.

Q6. What criteria determined which 47 methods were included, and how were their implementations checked against the source methods? Were any reproduced under their original settings? Five base skills score 0.0 or 0.5 mean ASR across all four development targets, and the reported gains are measured against these baselines.

Q7. Will the skills, execution histories, and evolution-agent prompt be released, and under what access conditions? Please also specify the refinement stopping rules, how canonical variants were chosen for the main experiments, and any human intervention during evolution.

Q8. Appendix B states that JailbreakSkill also represents attacks as executable skills and supports refinement, combination, and discovery, and locates the distinction in scope and initialization. Two things follow. First, beyond starting from a larger set of published methods and retaining their search and control flow, what methodological difference separates the two systems? Second, Appendix D reports that the Section 4.3 comparison matches the planner but retains JailbreakSkill's default auxiliary model, qwen-2.5-7b-instruct. Which auxiliary models do RedDarwin's skills use, and how does the comparison change when both systems are given the same one?

Q9. Do you plan to release the skill library, the evolved population, and the execution histories, and if so under what access conditions?

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

Related papers

Discussion (0)

Sign in to comment.