acceptodds
Under review as a conference paper at ICLR 2027

DARWIN: Evolving Jailbreak Adversary and Guardrail for LLM Safety Evaluation and Protection

Abstract

Most existing LLM safety evaluation and defense methods follow a static formulation: jailbreak vulnerabilities are evaluated with fixed attack methods, and guardrails are trained on fixed malicious prompt datasets. However, real-world adversaries continuously evolve their attack capabilities and expand the attack space. To address this challenge, we propose DARWIN, an evolutionary attack-defense framework that formulates jailbreaking as a continual evolutionary process and continuously updates guardrails through an attack-defense loop. We propose DARWIN-Attack as an evolutionary adversary that expands its attack capabilities through strategy discovery, mutation, and selection. DARWIN-Attack discovers new attack strategies from broad external sources, generates new variants through self-reflection and genetic evolution, and filters effective strategies according to their performance against aligned LLMs. During the attack execution phase, DARWIN-Attack adaptively selects and composes evolved strategies according to feedback from target LLMs and guardrails. Through continuous evolution, DARWIN-Attack achieves state-of-the-art attack success rates against frontier LLMs and guardrails, e.g., nearly 100% on DeepSeek-V4-Pro, over 90% on GPT-5.5 and nearly 100% on YuFeng-XGuard, outperforming recent evolving frameworks such as LSA and MAGIC. The continued evolution of DARWIN-Attack exposes new safety vulnerabilities and requires timely corresponding updates to safety defenses. Therefore, on the defense side, we introduce DARWIN-Guard, an online adversarial guardrail training paradigm, which iteratively trains the guardrail on the emerging adversarial samples generated by DARWIN-Attack. To improve robustness without sacrificing utility, DARWIN-Guard jointly learns from malicious and benign disguised queries, encouraging the guardrail to recognize underlying intent rather than superficial attack patterns. Through continuous evolution, DARWIN-Guard achieves an average unsafe recall of 95.0% across multiple safety evaluation benchmarks, outperforming recent advanced guardrails such as YuFeng and Nemotron. Meanwhile, DARWIN-Guard maintains an average benign pass rate of nearly 100% on standard benign benchmarks and achieves the best performance on over-refusal benchmarks. Our code and model are available at https://anonymous.4open.science/r/DARWIN-D8A6.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Evolving Jailbreak Attacks, LLM Safety Guardrails, Attack–Defense Co-evolution

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_2oi8OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

DARWIN couples an evolving black-box jailbreak attacker with an adversarially trained prompt guard.

  • DARWIN-Attack keeps an explicit strategy pool that grows from 50 seed strategies to 200. New strategies come from crawled external sources, genetic crossover and mutation, and failure reflection.
    • Candidates are deduplicated by embedding similarity. They are admitted only if they reach at least 0.8 ASR on a Qwen2.5-7B sandbox.
    • At attack time, strategies are composed up to three deep. The next strategy is sampled from a Markov transition matrix updated with a Q-learning-style rule.
    • The attacker and reflection model is Mistral-7B.
  • DARWIN-Guard fine-tunes Qwen3Guard-Gen-8B for 20 rounds. Training data is DARWIN-disguised harmful prompts (WildJailbreak) and benign prompts (OR-Bench), together with their raw versions.

Reported results:

  • The highest ASR of seven methods on HarmBench and AdvBench, across four frontier LLMs and two guards. Examples: 93.7% on GPT-5.5, 78.2% on Claude Sonnet 4.6, and about 99% on Qwen3Guard and YuFeng-XGuard. AQC is as low as 2.4.
  • For the guard: 95.0% average unsafe recall over nine benchmarks, about 100% benign pass on QA sets, and the best over-refusal scores on XSTest and JBB-Benign.
  • Co-evolution analyses (cross-stage ASR, held-out attack families) and ablations of each component and of the benign branch.
Critical strengths
  1. Benign disguising as the counterweight to adversarial guard training. Fig. 5 is the most informative result in the paper.

    • Removing benign disguising drops JBB-Benign pass rate from 80.0 to 59.0. Removing the benign branch drops it further to 25.0. Unsafe recall rises in both cases.
    • This shows that naive adversarial guard training mostly learns "jailbreak format ⇒ unsafe."
    • Symmetric disguising is a simple recipe that should transfer to other guards.
  2. Breadth of the attack evaluation. The paper covers:

    • current frontier targets and guards;
    • seven methods under a shared 60-query budget;
    • a second judge (App. A.7);
    • component ablations from a shared S50S_{50} checkpoint (Table 3).

    The margin on Claude (78.2% vs. 36.2% for MAGIC) is large, if it holds under the fairer comparisons requested below.

Critical weaknesses
  1. The guard is never evaluated against a strong adaptive attacker other than its own trainer.

    • Table 2 and Fig. 2 are static benchmarks, although the paper's own premise is that static evaluation underestimates vulnerability.
      • DARWIN-Attack against DARWIN-Guard is never reported as a Table 1 row.
      • The only adaptive number is in Fig. 3(B): Attack-200 vs. Guard-200 gives 30% ASR. So the final guard still misses almost a third of in-distribution attacks.
    • No baseline attack (TAP, MAJIC, the MAGIC or LSA attackers) is run against DARWIN-Guard.
      • The "held-out families" in Fig. 3(C) are produced by the same pipeline.
      • "Unseen" therefore means unseen strategies, not an unseen attacker.
    • The obvious data baseline is missing: Qwen3Guard fine-tuned on WildJailbreak's own adversarial_harmful and adversarial_benign splits.
      • These come from the same source dataset and already contain disguised benign prompts.
      • Static AT (Fig. 3A) uses DARWIN-generated data, so it does not isolate the value of the generator.
      • The gains on WildGuardTest and JailbreakHub may partly reflect how close those benchmarks are to the WildJailbreak training distribution.
    • The over-refusal evidence is thin.
      • The six QA sets are saturated at about 100% for every guard, so they say little.
      • OR-Bench is used for training, which leaves held-out over-refusal resting on XSTest-safe and JBB-Benign alone.
  2. The attack comparison is not apples-to-apples, and the gain isn't cleanly attributed to evolution.

    • The transition matrix and attack memory persist across test instances within each target–dataset pair (App. A.2).
      • DARWIN therefore adapts transductively over the benchmark, while TAP and ReNeLLM keep no cross-instance state.
      • A.2 lets baselines retain their own adaptive state, but this only benefits methods designed with cross-instance memory, so the comparison still favors DARWIN.
      • AQC of 2.4 on DeepSeek-V4-Pro and 2.7 on Qwen3Guard suggests convergence onto a few near-universal templates.
    • The pool is built from crawled community jailbreaks, with deduplication only against the pool itself.
      • A.1 excludes HarmBench/AdvBench prompts from strategy generation and admission. However, I could not find any check of the crawled templates themselves against benchmark items or known target-specific exploits.
      • "w/o External Knowledge" is the largest ablation drop (Claude 78.2 → 52.0). Much of the gain may come from the curated library rather than the evolutionary machinery.
    • The "Fixed" baseline in Fig. 4 samples strategies at random.
      • It removes Markov selection and reflection along with pool evolution.
      • It therefore does not show that evolution itself explains the 78.2 vs. 25.2 gap.
    • Novelty over MAJIC (Markov strategy composition) and AutoDAN-Turbo (lifelong strategy library) is incremental.
      • The update rule L1-normalizes TD-updated values into sampling probabilities.
      • This is ad hoc and is not analyzed.
    • All results are single runs with no variance reported. Several key gaps are small, e.g. 91.7 vs. 95.0 in Fig. 3A.
Questions for authors
  • What does Table 1 look like when the transition matrix and attack memory are reset for each instance? What fraction of successes come from the top-10 strategies?
  • What ASR do TAP, MAJIC, and the MAGIC attacker achieve against the final DARWIN-Guard under B = 60? Can the final DARWIN-Attack vs. DARWIN-Guard result be added as a Table 1 row?
  • How does a sample-matched Qwen3Guard SFT on WildJailbreak’s adversarial harmful and benign splits compare?
  • Is the 200-strategy pool in Table 1 the same pool that was co-evolved against DARWIN-Guard? How do the K-stages map onto the 20 guard training rounds? How were the external sources decontaminated against HarmBench/AdvBench and against known target-specific exploits?
6 comments
AuthorsOpenReview ↗

Response to Reviewer 2oi8 (1/2)

We thank Reviewer 2oi8 for the valuable comments. Below are our responses to the questions.

1. Adaptive Robustness and Generalization

We have evaluated DARWIN-Guard against strong adaptive baseline attackers on HarmBench and AdvBench with a budget of B=60 target queries per instance. DARWIN-Guard achieves the lowest average ASR among the evaluated models on both benchmarks. The results are reported in Appendix B.8 (Table 13). Furthermore, Figure 3(B) reports adaptive ASR for all attacker–guardrail checkpoint pairs across evolution stages, including 30% for Attack-200 against Guard-200.

The 30% ASR may mainly reflect the training–test distribution gap, rather than indicating that almost a third of in-distribution attacks are missed. Our harmful training data are from WildJailbreak, whereas our test data are from HarmBench. On harmful queries sampled from WildJailbreak’s test set, DARWIN-Guard achieves over 95% unsafe recall under the same DARWIN-Attack protocol (B=60), demonstrating effective in-distribution defense.

Although under DARWIN-Attack, DARWIN-Guard’s ASR on HarmBench is currently 30%, YuFeng-XGuard’s ASR is 99.2%. We have already observed a decrease in ASR from 52% to 30% as evolution progresses. We expect ASR to decrease further with continual defense evolution.

DatasetAttackGPT-5.5 ASR↓Claude Sonnet 4.6 ASR↓YuFeng-XGuard ASR↓DARWIN-Guard ASR↓
HarmBenchTAP35.5%1.5%44.2%1.7%
AutoDAN-Turbo48.7%11.2%64.7%10.5%
LSA58.5%21.2%73.5%18.2%
MAJIC68.5%30.7%82.7%24.5%
MAGIC73.5%36.2%80.5%23.2%
DARWIN-Attack93.7%78.2%99.2%30.0%
AdvBenchTAP32.5%1.1%41.7%1.1%
AutoDAN-Turbo45.1%9.6%61.1%8.4%
LSA55.5%18.6%70.3%16.7%
MAJIC65.1%27.5%79.4%22.3%
MAGIC69.8%32.6%77.3%21.1%
DARWIN-Attack90.7%68.2%96.7%27.5%

2. Comparison with WildJailbreak SFT

We added a sample-matched baseline (+ WJ SFT) by fine-tuning Qwen3Guard on WildJailbreak’s adversarial_harmful and adversarial_benign subsets under the same class ratio and training settings. Compared with Qwen3Guard, it yields limited recall gains on harmful benchmarks but lower benign pass rates (BPR) on XSTest and JBB-Benign. DARWIN-Guard achieves higher average unsafe recall and higher BPR:

BenchmarkQwen3Guard+ WJ SFTDARWIN-Guard
Aegis2.084.286.191.9
JBB-Behaviors98.099.0100.0
HarmBench98.298.599.0
S-Eval52.069.290.0
SemanticRouter74.876.887.6
OpenAI Moderation91.693.099.4
WildGuardTest84.887.291.8
StrongREJECT98.498.799.7
JailbreakHub80.483.895.2
Avg. Recall↑84.788.095.0
XSTest (BPR↑)95.284.897.6
JBB-Benign (BPR↑)62.041.080.0

3. Per-Instance Reset and Strategy Usage

We reset the transition matrix and attack memory before each input with all other settings unchanged. Under this per-instance reset, DARWIN-Attack (DARWIN(r)) still achieves higher ASR and lower AQC than all baselines:

DatasetTargetLSA ASR↑MAGIC ASR↑DARWIN(r) ASR↑LSA AQC↓MAGIC AQC↓DARWIN(r) AQC↓
HarmBenchDeepSeek-V4-Pro79.0%81.2%92.0%26.420.58.9
GPT-5.558.5%73.5%81.7%42.530.329.6
Gemini-3.5-Flash72.7%77.5%82.7%36.333.026.5
Claude Sonnet 4.621.2%36.2%62.5%57.852.838.7
Qwen3Guard81.5%82.5%92.7%20.417.39.3
YuFeng-XGuard73.5%80.5%90.5%32.727.314.8
AdvBenchDeepSeek-V4-Pro78.4%82.3%89.6%25.319.811.2
GPT-5.555.5%69.8%78.2%42.532.632.5
Gemini-3.5-Flash69.8%74.0%80.3%37.435.029.6
Claude Sonnet 4.618.6%32.6%54.8%57.051.643.1
Qwen3Guard79.2%79.6%91.5%23.521.912.1
YuFeng-XGuard70.3%77.3%87.6%34.830.018.2

For full DARWIN-Attack runs, Top-10 and Multi denote the percentages of successful inputs using any top-10 strategy (31.9–40.4%) and ≥2 distinct strategies (78.6–90.4%) on HarmBench (HB) and AdvBench (AB), respectively, showing diverse strategy use at low query cost:

TargetHB Top-10HB MultiAB Top-10AB Multi
DeepSeek-V4-Pro36.078.635.680.7
GPT-5.535.485.837.286.8
Gemini-3.5-Flash31.983.835.984.5
Claude Sonnet 4.637.089.440.290.4
Qwen3Guard35.880.440.483.2
YuFeng-XGuard32.984.636.386.6
AuthorsOpenReview ↗

Response to Reviewer 2oi8 (2/2)

4. Additional Over-Refusal Evaluation

We have added PHTest-harmless and CoCoNot-Contrast to our over-refusal evaluation. We report benign pass rate (BPR, %, higher is better). DARWIN-Guard achieves the highest BPR across all evaluated datasets. The results are shown below.

ModelXSTest BPR↑JBB-Benign BPR↑PHTest-harmless BPR↑CoCoNot-Contrast BPR↑
ShieldGemma80.879.088.990.8
Nemotron Guard77.665.077.088.1
Granite Guardian86.055.088.089.4
Llama Guard-396.877.089.793.1
Qwen3Guard95.262.080.491.6
YuFeng-XGuard94.059.085.992.3
DARWIN-Guard97.680.092.099.7

5. Strategy Pool Construction and Evolution

DARWIN-Attack expands its strategy pool from 50 to 200 through external knowledge evolution, genetic strategy evolution, and reflection-driven evolution, and adapts strategy selection and composition through feedback-guided evolution. New strategies are added only after similarity filtering and sandbox validation. The final 200-strategy pool is used in Table 1.

During evolution, we save Attack-K and Guard-K checkpoints at K ∈ {50, 75, 100, 125, 150, 175, 200} for evaluation, corresponding to the completion of guard training rounds {1, 3, 5, 8, 11, 15, 20}, respectively. The intervals between these rounds increase as the growing strategy pool makes it harder to find new strategies that satisfy both diversity and effectiveness requirements. Guard-50 is saved after round 1, while the initial guard corresponds to round 0.

We extract attack strategies rather than specific attack examples from external materials and use data from other datasets, not HarmBench or AdvBench, for strategy validation and co-evolution (Appendices A.1, A.3, and B.2). We also check source descriptions and model/version information to exclude exploits tailored to the evaluated model versions. We will clarify the checkpoint mapping and these procedures in the appendix.

6. Contributions of Evolutionary Components

Just as people improve by learning from external sources, DARWIN includes External Knowledge Evolution as one of its four evolutionary mechanisms. Table 3 evaluates each component from the same initial strategy pool. With external knowledge acquisition retained, genetic evolution, reflection, and Markov selection provide ASR gains of 21.7%, 19.2%, and 13.2%, respectively, on Claude Sonnet 4.6 compared with their corresponding ablated variants. These results demonstrate the additional benefits of strategy evolution and adaptive composition beyond external strategy collection. We will highlight these component contributions more clearly in the paper.

7. Novelty of DARWIN

(a) AutoDAN-Turbo learns and reuses strategies through attack-log summarization and retrieval. However, it lacks explicit mechanisms for genetic strategy evolution (e.g., crossover and mutation), rejection analysis to refine failed strategies, and feedback-driven adaptive updates to strategy transitions.

(b) MAJIC adaptively composes strategies from a fixed pool of 10 predefined strategies. Yet its pool does not evolve through external strategy acquisition. Existing strategies are neither evolved through crossover and mutation nor refined by analyzing rejection reasons.

In contrast, DARWIN-Attack jointly integrates four complementary evolutionary mechanisms: External Knowledge Evolution, Genetic Strategy Evolution, Reflection-Driven Evolution, and Feedback-Guided Evolution. It continuously evolves and expands its strategy pool, with 200 validated strategies, and adapts strategy selection and composition to target feedback without fine-tuning the attacker LLM. DARWIN also co-evolves an external guardrail through online adversarial training on harmful and benign disguised queries, enabling both sides to evolve while leaving the target LLM unchanged.

8. Stability Across Runs

We report mean ± standard deviation (%) over three runs with different random seeds for key guardrail and attack comparisons. The results show low variability and consistent advantages for DARWIN, including the comparison against w/o GDE in Figure 3(A). We will add these statistics to the revised paper.

Guardrail stability (Fig. 3(A)):

Variant in Fig. 3(A)Avg. unsafe recall↑
Static AT89.6 ± 0.5
Fixed Attacker (K=200)92.5 ± 0.4
w/o GDE91.7 ± 0.4
DARWIN95.0 ± 0.3

Attack stability:

DatasetMethodGPT-5.5 ASR↑Claude Sonnet 4.6 ASR↑
HarmBenchMAGIC73.5 ± 0.536.2 ± 0.7
DARWIN-Attack93.7 ± 0.478.2 ± 0.6
AdvBenchMAGIC69.8 ± 0.632.6 ± 0.8
DARWIN-Attack90.7 ± 0.868.2 ± 0.5

We hope our responses and revisions to the manuscript could address your concerns. We would be happy to answer any further questions.

Thank you for the additional experiments, particularly the comparisons against other baselines, the per-instance reset, and the smaller ask on strategy usage.

The per-instance reset results actually deepen my concern about the attack comparison, and I'd like the DARWIN(r) rows added to Table 1. The ASR drops under reset suggest that a meaningful part of the gain comes from state carried across test instances. That said, I appreciate how transparently the result is reported.

On a more minor note, I have questions regarding decontamination and the WJ SFT baseline. On decontamination, "we check source descriptions and model/version information" is not very informative. How many crawled sources were excluded, and on what basis? On the WJ SFT baseline, I suspect it is undertuned (JBB-Benign drops from 62 to 41), so I'd be curious about its training hyperparameters.

I have a similar concern about the external-sources analogy and Table 3. Both are quite vague, and they're rather hard to credit with only a single target (Claude Sonnet 4.6) and a single checkpoint.

I've raised my soundness score from 2 to 3 and am keeping my overall rating at 2.

Reviewer_QmcJOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

The paper presents a novel evolutionary framework for jailbreak attack and defense. An attack module iteratively retrieves strategies from external sources and develops new strategies via crossover and mutation. Strategies are only admitted after passing a capability filter. At attack time, attacks are composed via a Markov transition process that adapts over time, and feedback from failed attempts is used to refine attempted strategies. Attacks generated by DARWIN-Attack are used to train a safeguard LLM, which learns to classify harmful queries based on both harmful and benign requests embedded within diverse attack strategies. A thorough evaluation and set of ablations demonstrate the efficacy of both modules.

Critical strengths
  • The paper introduces a new approach to evolutionary jailbreak algorithms that stores and evolves attack strategies externally instead of within model weights. The added control and explicit flexibility this provides seems to significantly improve performance, as demonstrated by the evaluations, and I think this finding could be a meaningful contribution to the field.
  • The evaluation and ablations are thorough, and the additional analysis they provide regarding which components are impactful is interesting.
Critical weaknesses
  • The paper's methods section is somewhat difficult to follow and lacks key technical details, even after reading through the appendix. How is external retrieval actually implemented? (i.e. is this done autonomously by an agent via web search? another way?). How are mutation operations actually carried out? How are crossover operations carried out? In what format are strategies stored? Are attack prompts generated by an LLM or by embeddings strings into predefined templates? Is there any prioritization of what is added to the list of strategies beyond passing the success threshold? Which backend models were used for each baseline? Figure 1 is also very dense, making it difficult to understand when each step occurs. Without a clear presentation of the employed methods, it is difficult to properly evaluate the work.
  • The paper's framing makes it difficult to understand how the paper intends to relate itself to prior work. Which components of the proposed strategy are novel versus taken from prior methods? How specifically does this work intend to improve upon prior evolutionary jailbreak algorithms and why? Additionally, descriptive and quantitative comparison against the relevant baseline AutoRedTeamer (Zhao et al., 2025) is missing.
  • If possible, the trained safeguard model should be evaluated against the baseline jailbreak attacks to determine whether or not the performance gains generalize to other strong attack strategies.
Questions for authors

Please see the weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer QmcJ (1/2)

We thank Reviewer QmcJ for the valuable comments. Below are our responses to the questions.

1. Technical Details and Execution Order

ComponentImplementation details
External knowledgeAn agent searches public sources, including technical documentation, community discussions, and institutional safety reports. The retrieved materials are then analyzed by an unaligned LLM such as Mistral-7B-Instruct-v0.2 to extract reusable strategy instructions.
Mutation and crossoverMutation uses an LLM to rewrite a strategy with one of 15 operators. Crossover uses the LLM to combine two parents sampled from the top-5. Parents are ranked by historical attack ASR, or sandbox ASR if history is unavailable. Each operation has probability 0.5.
Strategy storageJSONL records store IDs, names, instructions, and metadata; SQLite stores performance statistics, embeddings, and transition values.
Prompt generationThe attacker LLM rewrites the current prompt following the selected strategy while preserving intent. Strategies are applied sequentially for composition.
Admission and prioritizationCandidates pass semantic deduplication (cosine threshold 0.80) and sandbox validation (ASR≥80% on Qwen2.5-7B-Instruct), then are admitted in processing order until the pool is full. Performance ranking is used for parent selection.
Baseline backendsTAP, ReNeLLM, AutoDAN-Turbo, MAJIC and AutoRedTeamer use Mistral-7B-Instruct-v0.2, as does DARWIN-Attack. LSA and MAGIC start from their authors’ released attacker models because they evolve attackers through parameter updates.

Each evolution round updates the strategy pool using external knowledge, genetic operations, and failure feedback, then generates harmful and benign disguised queries against the current guardrail. Target feedback guides prompt refinement and subsequent strategy selection and composition. Disguised and original queries train the guardrail for the next round. During final attack evaluation, selection, composition, and prompt refinement adapt to target feedback. We have clarified the implementation details and execution order in Appendix A.7 (Table 6) and refined Figure 1’s stage numbering and feedback arrows. Our code and trained guardrail are released for implementation details and reproducibility.

AuthorsOpenReview ↗

Response to Reviewer QmcJ (2/2)

2. Contributions and Comparison with AutoRedTeamer

(a) AutoDAN-Turbo learns and reuses strategies through attack-log summarization and response-similarity retrieval. However, it lacks explicit mechanisms for genetic strategy evolution (e.g., crossover and mutation), rejection analysis to refine failed strategies, and feedback-driven adaptive updates to strategy transitions.

(b) MAJIC adaptively selects and composes strategies from a fixed pool of 10 predefined strategies. Yet its pool does not evolve through external strategy acquisition. Existing strategies are neither evolved through genetic operations such as crossover and mutation nor refined by analyzing rejection reasons.

(c) AutoRedTeamer (Zhou et al., 2025) discovers attacks from the literature and uses Attack Memory to select and combine them. However, it lacks explicit strategy-level genetic evolution (e.g., crossover and mutation) and does not analyze rejection reasons to refine failed strategies. Its framework also lacks an explicit feedback-driven update rule for adaptively adjusting strategy transitions.

In contrast, DARWIN-Attack jointly integrates four evolutionary mechanisms: External Knowledge Evolution, Genetic Strategy Evolution, Reflection-Driven Evolution, and Feedback-Guided Evolution (Section 3.2), continuously evolving its strategy pool and adapting strategy selection and composition. Unlike LSA and MAGIC, it does not fine-tune the attacker LLM. DARWIN also co-evolves an external guardrail through online adversarial training on harmful and benign disguised queries to improve robustness and mitigate over-refusal.

Table 3 quantifies the contributions of DARWIN's evolutionary mechanisms. On Claude Sonnet 4.6, removing external knowledge, genetic evolution, reflection, and Markov selection individually reduces ASR from 78.2% to 52.0%, 56.5%, 59.0%, and 65.0%, respectively. We also evaluated AutoRedTeamer under the same experimental settings described in Appendices A.1–A.2. DARWIN-Attack achieves higher ASR and lower AQC on all evaluated targets and higher HS on all evaluated LLMs. We have clarified these contributions and comparisons in Sections 2.1 and 4.5 and included the AutoRedTeamer results in Table 1. The results are shown below.

DatasetTargetAR ASR↑DARWIN ASR↑AR HS↑DARWIN HS↑AR AQC↓DARWIN AQC↓
HarmBenchDeepSeek-V4-Pro78.2%99.7%0.380.6930.42.4
HarmBenchGPT-5.552.0%93.7%0.230.5347.618.4
HarmBenchGemini-3.5-Flash69.0%93.0%0.260.5538.715.9
HarmBenchClaude Sonnet 4.614.5%78.2%0.090.4758.628.2
HarmBenchQwen3Guard82.2%99.7%——20.92.7
HarmBenchYuFeng-XGuard68.0%99.2%——36.13.7
AdvBenchDeepSeek-V4-Pro75.0%97.6%0.350.6530.64.5
AdvBenchGPT-5.548.2%90.7%0.200.5145.920.5
AdvBenchGemini-3.5-Flash66.1%90.9%0.240.5239.821.0
AdvBenchClaude Sonnet 4.612.5%68.2%0.080.4158.534.9
AdvBenchQwen3Guard80.1%98.8%——24.65.8
AdvBenchYuFeng-XGuard64.4%96.7%——38.58.9

Note: AR denotes AutoRedTeamer; DARWIN denotes DARWIN-Attack; "—" indicates HS is not applicable to guardrails.

3. Generalization to Other Jailbreak Attacks

We have evaluated DARWIN-Guard against strong baseline jailbreak attacks on HarmBench and AdvBench using the same experimental settings described in Appendices A.1–A.2. DARWIN-Guard achieves the lowest ASR among the evaluated models on both benchmarks, supporting its generalization to other strong attack strategies. We have included these additional experimental results in Appendix B.8 (Table 13) and summarized them in Section 4.3. The results are shown below.

DatasetAttackGPT-5.5 ASR↓Claude Sonnet 4.6 ASR↓YuFeng-XGuard ASR↓DARWIN-Guard ASR↓
HarmBenchAutoDAN-Turbo48.7%11.2%64.7%10.5%
HarmBenchLSA58.5%21.2%73.5%18.2%
HarmBenchMAJIC68.5%30.7%82.7%24.5%
HarmBenchMAGIC73.5%36.2%80.5%23.2%
AdvBenchAutoDAN-Turbo45.1%9.6%61.1%8.4%
AdvBenchLSA55.5%18.6%70.3%16.7%
AdvBenchMAJIC65.1%27.5%79.4%22.3%
AdvBenchMAGIC69.8%32.6%77.3%21.1%

We hope our responses and revisions to the manuscript could address your concerns. We would be happy to answer any further questions.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer QmcJ,

Thank you again for your thoughtful and constructive feedback, which we have carefully addressed in our author response. As the discussion deadline is approaching, we would greatly appreciate any further feedback you may have. If our responses have resolved your main concerns, we would also be grateful if you could update your overall assessment accordingly.

Sincerely,

The Authors

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

This paper proposes DARWIN, a framework in which an attacker and a guardrail co-evolve. DARWIN-Attack maintains an explicit strategy pool that grows through external knowledge, genetic evolution (crossover and mutation), and reflection on failed attempts, and it uses target feedback to select and compose strategies. DARWIN-Guard is trained through online adversarial training on both harmful and benign prompts disguised by DARWIN-Attack, so it learns to recognize underlying intent rather than surface attack patterns.

Critical strengths
  • The results are strong. DARWIN-Attack clearly outperforms prior methods, including recent co-evolving frameworks such as LSA and MAGIC, and DARWIN-Guard achieves the highest average unsafe recall while keeping over-refusal low.
  • I find the attacker design a notable contribution. Building an explicit, expanding strategy pool and evolving it with genetic operators, without fine-tuning the attacker LLM, is a nice way to keep previously effective strategies and avoid catastrophic forgetting.
  • The experiments are thorough. They cover multiple target LLMs and guardrails, benign and over-refusal benchmarks, cross-stage and held-out attack-family analyses, and ablations of each evolutionary component.
Critical weaknesses
  • Presenting the shift from a static paradigm to a continuous attack–defense loop as a primary contribution feels somewhat overclaimed. The authors do discuss LSA (Wang et al., 2025) and MAGIC (Wen et al., 2026) in the related work, but the introduction and contribution list still describe “existing works” as static. I would suggest positioning the contribution more precisely, for example around the explicit evolving strategy pool, adaptive composition, and the external guardrail, rather than around the loop itself.
    • The use of a strategy pool is also not entirely new. AutoDAN-Turbo also builds reusable jailbreak strategies, and MAJIC already composes strategies through a Markov chain. It would help to clarify more explicitly how DARWIN-Attack differs beyond expanding and evolving the pool, and which part of the gain comes from this difference.
  • The target models (DeepSeek-V4-Pro, GPT-5.5, Gemini-3.5-Flash, and Claude Sonnet 4.6) are a reasonable set, but it would strengthen the paper to include stronger frontier models such as GPT-5.6 or Opus 5, since they are available in August. This would show whether the attack remains effective against the most capable and best-aligned models.
Questions for authors

na

1 comment
AuthorsOpenReview ↗

Response to Reviewer idCP

We thank Reviewer idCP for the valuable comments. Below are our responses to the questions.

1. Positioning of Our Contributions

We would like to clarify DARWIN's technical contributions on both the attack and defense sides. On the attack side, DARWIN integrates four complementary evolutionary mechanisms, including External Knowledge Evolution, Genetic Strategy Evolution, Reflection-Driven Evolution, and Feedback-Guided Evolution. These mechanisms enable the acquisition, refinement, and adaptive composition of reusable attack strategies without fine-tuning the attacker LLM. On the defense side, DARWIN employs online adversarial training of an external guardrail using both harmful and benign disguised queries to improve jailbreak robustness and mitigate over-refusal. When describing existing work as static, we use “most” in the abstract and “typically” at the beginning of the introduction, and discuss existing co-evolving methods such as LSA and MAGIC in Section 2.2. Following your suggestion, we will further refine the introduction and contribution list to better distinguish DARWIN from prior work and highlight its specific technical contributions.

2. Differences from AutoDAN-Turbo and MAJIC

(a) AutoDAN-Turbo learns and reuses strategies through attack-log summarization and response-similarity retrieval. However, it lacks explicit mechanisms for genetic strategy evolution (e.g., crossover and mutation), rejection analysis to refine failed strategies, and feedback-driven adaptive updates to strategy transitions.

(b) MAJIC adaptively selects and composes strategies from a fixed pool of 10 predefined strategies. Yet its pool does not evolve through external strategy acquisition. Existing strategies are neither evolved through genetic operations such as crossover and mutation nor refined by analyzing rejection reasons.

(c) DARWIN-Attack integrates four evolutionary mechanisms: External Knowledge Evolution, Genetic Strategy Evolution, Reflection-Driven Evolution, and Feedback-Guided Evolution (Section 3.2). It continuously evolves and expands its strategy pool by the above four evolution mechanisms, resulting in 200 validated strategies (the largest strategy pool among the compared methods). Feedback from the evolving guardrail can further guide adaptive strategy selection and composition.

Table 3 quantifies the contributions of DARWIN's evolutionary mechanisms to attack effectiveness and query efficiency. On Claude Sonnet 4.6, removing external knowledge, genetic evolution, reflection, and Markov selection individually reduces ASR from 78.2% to 52.0%, 56.5%, 59.0%, and 65.0%, respectively. Removing Markov selection also increases the average query count on GPT-5.5 from 18.4 to 39.0.

3. Evaluation on Stronger Frontier Models

We have added evaluations on GPT-5.6 and Opus 5 using the same experimental settings as in the paper. DARWIN-Attack achieves the highest ASR and HS and the lowest AQC among the evaluated methods on both HarmBench and AdvBench. The results are shown below.

DatasetMethodGPT-5.6 ASR↑HS↑AQC↓Opus 5 ASR↑HS↑AQC↓
HarmBenchAutoDAN-Turbo37.7%0.1453.42.2%0.0259.8
HarmBenchLSA46.5%0.2047.410.7%0.0659.0
HarmBenchMAJIC56.0%0.2740.519.2%0.1157.8
HarmBenchMAGIC61.0%0.3036.825.0%0.1455.8
HarmBenchDARWIN-Attack87.2%0.4923.472.7%0.4332.0
AdvBenchAutoDAN-Turbo33.6%0.1251.81.1%0.0159.9
AdvBenchLSA43.0%0.1847.08.2%0.0558.8
AdvBenchMAJIC53.0%0.2442.216.5%0.0956.8
AdvBenchMAGIC58.2%0.2738.521.5%0.1254.8
AdvBenchDARWIN-Attack83.8%0.4625.762.6%0.3738.2

We hope our responses and revisions to the manuscript could address your concerns. We would be happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.