- 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
-
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.
-
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 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
-
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_harmfulandadversarial_benignsplits.- 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.
- Table 2 and Fig. 2 are static benchmarks, although the paper's own premise is that static evaluation underestimates vulnerability.
-
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.
- The transition matrix and attack memory persist across test instances within each target–dataset pair (App. A.2).
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
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.
| Model | XSTest BPR↑ | JBB-Benign BPR↑ | PHTest-harmless BPR↑ | CoCoNot-Contrast BPR↑ |
|---|---|---|---|---|
| ShieldGemma | 80.8 | 79.0 | 88.9 | 90.8 |
| Nemotron Guard | 77.6 | 65.0 | 77.0 | 88.1 |
| Granite Guardian | 86.0 | 55.0 | 88.0 | 89.4 |
| Llama Guard-3 | 96.8 | 77.0 | 89.7 | 93.1 |
| Qwen3Guard | 95.2 | 62.0 | 80.4 | 91.6 |
| YuFeng-XGuard | 94.0 | 59.0 | 85.9 | 92.3 |
| DARWIN-Guard | 97.6 | 80.0 | 92.0 | 99.7 |
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 AT | 89.6 ± 0.5 |
| Fixed Attacker (K=200) | 92.5 ± 0.4 |
| w/o GDE | 91.7 ± 0.4 |
| DARWIN | 95.0 ± 0.3 |
Attack stability:
| Dataset | Method | GPT-5.5 ASR↑ | Claude Sonnet 4.6 ASR↑ |
|---|---|---|---|
| HarmBench | MAGIC | 73.5 ± 0.5 | 36.2 ± 0.7 |
| DARWIN-Attack | 93.7 ± 0.4 | 78.2 ± 0.6 | |
| AdvBench | MAGIC | 69.8 ± 0.6 | 32.6 ± 0.8 |
| DARWIN-Attack | 90.7 ± 0.8 | 68.2 ± 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.
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.
2. Comparison with WildJailbreak SFT
We added a sample-matched baseline (
+ WJ SFT) by fine-tuning Qwen3Guard on WildJailbreak’sadversarial_harmfulandadversarial_benignsubsets 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: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: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: