acceptodds
Under review as a conference paper at ICLR 2027

DART: Defense-Aware Red Teaming with a Critical Cue Bank for Open-World Adversarial Evaluation

Abstract

Autonomous red-teaming agents can inspect source code, invoke attack libraries, retrieve external knowledge, and use execution feedback, yet these capabilities do not ensure that the agent can identify relevant knowledge and operationalize it as an executable attack tailored to a particular defense. We introduce DART, a defense-aware open-world red-teaming framework centered on a Critical Cue Bank (CCB). The CCB structures profiling evidence into actionable cues that guide toolbox attack adaptation, external retrieval, and research-to-executable attack construction. When no directly applicable attack is available, DART retrieves strategies developed for related defense mechanisms and adapts their principles to the target's interface and threat-model constraints. Across 51 published defense implementations from AutoAdvExBench, DART increases the overall attack success rate from 5.9% to 43.1% with Qwen and from 31.4% to 52.9% with Claude Opus, while producing valid evaluations for all 51 targets. It also reduces LLM token usage by 84-95% relative to the corresponding baselines. An open-world baseline without cue-conditioned selection and adaptation yields little or no improvement in attack success rate, suggesting that the main bottleneck is not access to attack knowledge but its target-conditioned operationalization.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Autonomous Red Teaming, Adversarial Robustness Evaluation, Adaptive Attacks, LLM Agents, AI Safety Evaluation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WWgrOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

the paper introduces DART, a defense-aware open-world red-teaming framework centered on a Critical Cue Bank (CCB). DART first profiles the target defense and structures the resulting attack-relevant characteristics in the CCB, which guides attack selection, open-world retrieval, and target-specific adaptation.

Critical strengths
  • topic is very relevant and the paper is suitable for this conference
  • the goal of the paper is very clear, and the writing is clear and effective
  • the result of the paper is very strong and convincing
Critical weaknesses
  • currently only evaluated in one source of benchmark, not sure about broad applicability
  • no error bars shown in the results
Questions for authors

N/A

1 comment
AuthorsOpenReview ↗

Response to Reviewer WWgr: Generalization and Evaluation Stability (CW1–2)

Reviewer Concerns Addressed — Critical Weaknesses 1–2

CW1 — Generalization: The reviewer notes that DART was evaluated on a single benchmark and questions whether its effectiveness extends to other defense implementations and datasets.

CW2 — Evaluation Stability: The reviewer highlights the absence of error bars and run-to-run variability estimates, raising concerns about the reliability of the reported results.

Our Response

We thank Reviewer WWgr for the positive assessment and constructive suggestions. We conducted additional experiments on (1) external generalization and cross-dataset adaptation and (2) run-to-run variability and evaluation-set stability. Detailed results are provided in Shared Responses (3/4)–(4/4), supplemented by the cross-dataset experiments below.

1. Generalization Beyond AutoAdvExBench (CW1)

In Shared Response (3/4), we evaluate five external defenses across MNIST, CIFAR-10, GTSRB, and ImageNet. DART completes 5/5 evaluations (zero Failures) and achieves 3/5 Successes (60%):

  • TopoReformer: APGD-DLR; RA 0.0%.
  • ZeroPur: BPDA + CW-Margin PGD; RA 9.5%.
  • LeapNet-1: EOT + Margin-PGD; RA 28.5%.

For LeapNet-1, DART identifies randomized execution and adapts an EOT-based attack, reducing RA from 55.0% to 28.5% (−26.5 pp) and changing Not Broken to Success.

DART also generates executable defense-adapted attacks for GradNorm/Swin-B and RMAggNet, obtaining valid final RAs of 56.5% and 73.0%, respectively. Both remain Not Broken rather than failing during execution.

Cross-Dataset Evaluation with Retrained Defenses

To further examine generalization, we retrained TRADES, kWTA-Activation, and Mixup-Inference on MNIST while preserving their defense mechanisms. We evaluated 200 images per defense using MNIST ε = 0.3 and the original CIFAR-10 attack constraints.

DefenseCIFAR-10 Clean / RA (%)MNIST Clean / RA (%)Outcome
TRADES86.0 / 54.5100.0 / 92.5Not Broken in both
kWTA-Activation85.0 / 4.099.5 / 0.0Success in both
Mixup-Inference95.5 / 4.0100.0 / 0.0Success in both

DART uses gradient-based PGD for TRADES, surrogate gradients for discontinuous kWTA activations, and EOT for randomized Mixup-Inference.

All three defenses retain their Success/Not Broken outcomes after dataset change and retraining.

The external evaluations demonstrate applicability beyond AutoAdvExBench, while the retraining experiments provide complementary evidence of mechanism-aware adaptation across datasets. Absolute RA values are not directly comparable across different datasets and perturbation budgets.

2. Error Bars and Evaluation Stability (CW2)

In Shared Response (4/4), we report independent full-agent repetitions and evaluations using larger image sets.

Independent Agent Runs

Using Claude Opus, we independently repeated the complete DART workflow three times on each of five defenses, restarting profiling, attack generation, and final RA scoring.

DefenseMean RA ± Sample SD
TRADES54.50 ± 0.00
kWTA-Activation4.00 ± 3.12
Mixup-Inference4.00 ± 2.18
adversarial_detector0.33 ± 0.29
DiffPure88.50 ± 1.00

RA is in percent; sample SD is in percentage points.

All 15/15 independent executions produced valid RA, and all five defenses retained consistent Success/Not Broken outcomes across runs. Even kWTA-Activation, with the largest RA variation (6.0 pp), remained Success in all three runs.

Larger Evaluation Set

We additionally evaluated final attacks on 1,000 instead of 200 images across all 51 defenses.

Both settings yielded 27 Success and 24 Not Broken outcomes, with zero classification changes. The mean absolute RA difference was 1.00 pp, with a median of 0.60 pp and maximum of 6.10 pp.

For the near-threshold defense obfuscated-gradients/sap, RA changed from 44.0% (200 images) to 44.8% (1,000 images), while remaining Not Broken.

These experiments examine distinct sources of stability: independent agent-run variability on five defenses and evaluation-set-size sensitivity across all 51 defenses. The 200-image subset is included in the 1,000-image set.

3. Manuscript Revision

We will incorporate the external and cross-dataset evaluations, per-run RA and sample standard deviations, expanded evaluation results, and near-threshold analysis into the revised manuscript.

Together, these findings address CW1 and CW2 by providing evidence of applicability across heterogeneous defenses and consistent evaluation outcomes under independent agent executions and larger evaluation sets.

We thank Reviewer WWgr again for the constructive suggestions, which helped strengthen DART's experimental validation.

Reviewer_7vzBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents DART, an autonomous framework for constructing adaptive adversarial attacks against image-classification defenses. It profiles each target implementation, organizes attack-relevant evidence into eight Critical Cue Bank (CCB) categories, and uses these cues to guide both toolbox attack adaptation and retrieval of external attack papers or code. On AutoAdvExBench defense implementations, DART is shown to increase defense-level attack success considerably with Qwen and Claude Opus, while producing valid evaluations for all targets. DART is also shown to reduce LLM token use substantially.

Critical strengths
  • The paper presents a structured workflow for turning target observations and existing attack knowledge into executable attacks.
  • The poaper shows substantial practical gains across heterogeneous implementations and two backbones from using their approach.
  • The paper includes a repair ablation and detailed case analyses that shows concrete analyses to connect diagnostic evidence to attack construction and acknowledge failures. These analyses help explain how the workflow operates and why retrieving relevant knowledge can still be insufficient.
Critical weaknesses
  • The role of the central CCB needs a more direct ablation. Baseline + Open-World tests adding retrieval to the original agent, but DART also changes profiling, attack-side interface construction, candidate selection, and the execution/repair workflow. There is no otherwise identical DART variant using unstructured profiling observations instead of the cues. The repair ablation in Appendix B is useful but does not resolve this attribution problem. More evidence is needed to establish how much the CCB contributes to the overall gains.

  • The evaluation does not sufficiently separate improvements in attack strength from improved execution coverage and differences in search resources. The headline success rates count all 51 targets, including many baseline execution failures. Paired robust accuracies and outcome transitions on targets evaluated by both methods would help assess attack-strength gains separately. Aggregate query counts and normalization by completed evaluations do not establish comparable search budgets per target.

  • The paper specifies DART’s cap of five research candidates and its repair policy, but provides insufficient detail about baseline timeouts, stopping rules, profiling budgets, and query counting to assess budget comparability.

  • All configurations use one 200-image subset, with no reported repeated agent runs or systematic validation on a larger evaluation set. The stability of results near the success threshold remains unclear.

Questions for authors
  1. Can you report an otherwise comparable DART variant that passes the same profiling observations as an unstructured summary instead of the CCB, keeping the remaining prompts, retrieval tools, candidate limits, and execution budgets unchanged?
  2. Does the query metric include diagnostic probes, surrogate-model calls, backward passes, and repeated EOT samples? How are batched inputs counted?
  3. Are attack-side forward-pass construction and environment setup equally available to the baselines? Can you provide paired outcomes and robust accuracies on targets evaluated by both methods?
  4. How was the 200-image subset selected? Do the main outcome changes persist across additional agent runs and a larger evaluation set?
8 comments
AuthorsOpenReview ↗

Response to Reviewer 7vzB (1/6): CCB Ablation and Controlled Profiling Comparison

Reviewer Concerns Addressed — Critical Weakness 1 / Question 1

The reviewer notes that DART changes multiple components beyond the CCB, making its specific contribution difficult to isolate. In Question 1, the reviewer requests an otherwise identical DART variant that replaces structured CCB cues with unstructured profiling observations, while preserving the remaining prompts, retrieval tools, candidate limits, and execution budgets.

Our Response

We thank Reviewer 7vzB for emphasizing the need to isolate the CCB's contribution. We conducted two complementary experiments: (1) CCB-utilization ablations across ART and OpenWorld, and (2) a representation-controlled comparison using identical profiling observations. Detailed results are provided in Shared Responses (1/4) and (2/4), respectively.

1. CCB Utilization Ablation

Using Claude Opus, we evaluated four configurations on the same 51 defenses.

ConfigurationAttack SuccessCompleted RA
No CCB13/51 (25.5%)34/51
ART-only CCB22/51 (43.1%)51/51
OpenWorld-only CCB24/51 (47.1%)50/51
Full DART27/51 (52.9%)51/51

Extending CCB utilization from ART-only to both stages increased Success from 22 to 27 defenses and decreased mean paired RA from 44.16% to 35.60% (−8.56 pp). Both configurations completed all 51 evaluations, separating this attack-strength improvement from evaluation coverage.

2. Structured CCB vs. Flattened Observations

To directly address Question 1, we held the diagnostic measurements, backbone, remaining prompts, retrieval tools, candidate limits, attack procedure, and configured budgets constant.

The Flattened variant removed the explicit hierarchy of eight CCB categories while preserving the diagnostic field names and values.

MetricStructuredFlattened
Attack Success27/5127/51
Completed final RA51/5148/51
Mean RA on common 4833.05%32.83%
Total runtime36.62 h50.5 h
LLM API costUSD 99.46USD 119.63

Both representations broke the same 27 defenses. On the 48 jointly completed targets, Structured achieved lower RA on 16, Flattened on 17, and 15 were tied.

Therefore, the hierarchy alone did not improve attack strength on the jointly completed defenses. Its observed benefits concerned evaluation completion and resource use: Structured completed three additional evaluations and incurred lower runtime and API cost.

These experiments distinguish the contribution of utilizing defense-specific diagnostic information from that of organizing the same information into structured CCB categories. We will make this distinction explicit in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 7vzB (2/6): Target-Query Definition and Accounting

Reviewer Concerns Addressed — Critical Weakness 3 / Question 2

The reviewer requests precise query-counting rules to assess evaluation budgets. Specifically, Question 2 asks whether diagnostic probes, surrogate-model calls, backward passes, and repeated EOT samples are included, and how batched inputs are counted.

Our Response

We thank Reviewer 7vzB for requesting a precise and reproducible definition. We clarify the reported Target Query metric, specify which operations it includes, and provide an example from the evaluation logs. Aggregate LLM usage, runtime, and API costs are reported in Response (3/6), while search-budget settings are detailed in Response (5/6).

1. Definition of Target Queries

Target Queries count the total number of images evaluated during recorded robust accuracy (RA) scoring events on the original defense.

Let MM be the number of recorded scoring events and njn_j the number of images evaluated in event jj. The query total is:

Q=∑j=1MnjQ = \sum_{j=1}^{M} n_j

For example, scoring 200 images three times produces 600 Target Queries. Re-evaluating the same images with another candidate counts them again.

2. Operation-Level Accounting

OperationIncluded in Target Queries?
Original-defense RA scoringYes, per evaluated image
Diagnostic probesOnly when recorded as original-defense RA scoring
Other diagnostic model evaluationsNo
Surrogate-model forward callsNo
Backward passes / autogradNo
Internal EOT forward/backward samplesNo
Original-defense RA scoring of EOT attacksYes, per evaluated image
Batched RA evaluationCount evaluated images, not batch calls
Baseline final submission/checkAccounted for separately

Thus, Target Queries measure recorded original-defense RA-scoring activity, rather than every forward or backward operation performed during attack optimization.

3. Example from Actual Logs

For DiffPure using Claude Opus, we recorded:

  • Two RA-scoring events evaluating 64 images each.
  • Six RA-scoring events evaluating 200 images each.

Therefore, eight scoring events correspond to:

Q=2×64+6×200=1,328Q = 2 \times 64 + 6 \times 200 = 1{,}328

The batch-call count is eight, whereas the image-level Target Query count is 1,328.

4. Aggregate Accounting

Applying the same rules to the Claude Opus DART evaluation logs across 51 defenses yielded 580 recorded RA-scoring events and 96,400 Target Queries. The corresponding Claude Opus Baseline total was 38,600 Target Queries.

We will include the metric definition, inclusion and exclusion rules, batch-counting procedure, and representative log examples in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 7vzB (3/6): Resource Costs and Computational Analysis

Response to Reviewer 7vzB (3/6): Resource Costs and Computational Analysis

Reviewer Concerns Addressed — Critical Weaknesses 2–3 / Question 2

The reviewer requests clearer separation of attack effectiveness, evaluation coverage, and resource usage, along with detailed budget information. Target Query accounting is addressed in Response (2/6).

Our Response

We thank Reviewer 7vzB for this suggestion. We report LLM usage, Target Queries, runtime, API costs, and evaluation coverage separately.

1. Baseline–DART Resource Comparison

MetricQwen BaselineQwen DARTOpus BaselineOpus DART
LLM calls3,1377484,955570
LLM tokens44.48M7.22M167.87M8.35M
Target Queries81,200147,00038,60096,400
Runtime (h)44.2635.1876.9036.62
Completed RA16/5151/5122/5151/51

DART completes all 51 evaluations with both backbones, versus 16 (Qwen) and 22 (Opus) for Baseline.

With Claude Opus, DART reduces LLM calls by 88.5%, tokens by 95.0%, and aggregate runtime by 52.4%. API costs decrease from USD 158.27 to USD 99.46 (−37.2%).

Total Target Queries increase from 38,600 to 96,400 with Opus, alongside expanded evaluation coverage from 22 to 51 defenses.

Reported runtime is the sum of per-defense execution times.

2. Resources per Completed Evaluation

We normalize aggregate resource consumption by the number of completed evaluations.

MetricBaselineDARTChange
Qwen Target Queries5,0752,882−43.2%
Qwen runtime (h)2.770.69−75.1%
Opus Target Queries1,7551,890+7.7%
Opus runtime (h)3.500.72−79.5%
Opus API cost (USD)7.191.95−72.9%

For Qwen, DART reduces Target Queries per completion by 43.2% and runtime by 75.1%.

For Claude Opus, runtime per completion decreases by 79.5% and API cost by 72.9%, while Target Queries increase by 7.7%.

These results highlight DART's lower LLM overhead and execution time per completed evaluation. Further details on paired attack outcomes and search budgets are provided in Responses (4/6) and (5/6), respectively.

3. Target Query Accounting and Efficiency

As defined in Response (2/6), Target Queries count images evaluated during recorded robust-accuracy scoring on the original defense. Surrogate forward passes, backward passes, and internal EOT computations are not included in this metric.

Thus, Target Queries represent recorded RA-scoring activity, not total attack-optimization computation.

DART integrates defense profiling, CCB-guided attack selection, target-specific adaptation, and execution-feedback refinement. The results show that this workflow requires substantially fewer LLM interactions while completing more evaluations.

Execution limits, candidate budgets, repair policies, and supplementary search-budget experiments are detailed in Response (5/6).

4. Summary and Manuscript Revision

DART achieves 51/51 evaluation coverage while substantially reducing LLM calls, tokens, and reported runtime across both backbones, as well as API costs with Claude Opus.

These findings support DART's contribution to autonomous, defense-aware construction of executable adaptive attacks, particularly in evaluation coverage and LLM-side efficiency.

We will incorporate the resource comparisons, Target Query definitions, normalized costs, and execution-budget details into the revised manuscript, distinguishing LLM efficiency from attack-side computational requirements.

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

This paper proposes DART, an LLM-based framework for automatically generating adaptive attacks against adversarial defenses. Its main component, the Critical Cue Bank (CCB), collects information about a target defense and uses it to guide attack selection, external knowledge retrieval, and attack implementation.The authors evaluate DART on 51 defenses from AutoAdvExBench using two LLM backbones. The results show clear improvements in attack success rate and evaluation coverage, along with substantially lower LLM token usage compared with the baselines.

Critical strengths
  1. The paper addresses an important problem of automatically constructing adaptive attacks against real-world defenses.Using defense-specific information to guide attack selection and adaptation is well-motivated, and the Critical Cue Bank provides a structured way to support this process.
  2. Experiments on AutoAdvExBench show substantial improvements in attack success rates and evaluation coverage over the baselines, while significantly reducing LLM token usage and runtime.
Critical weaknesses
  1. The framework mainly combines existing techniques such as defense profiling, attack selection, and external knowledge retrieval. While the overall approach is effective, the specific benefit of the CCB is not clearly demonstrated. It remains unclear whether providing the agent with the same profiling information in plain text, without the structured cues, would achieve similar results.
  2. The evaluation is largely limited to a single benchmark(AutoAdvExBench) and its open-world variant, without considering other automated attack methods or alternative agent designs. It remains unclear how well the framework generalizes to unseen defenses or different evaluation settings.
Questions for authors
  1. Could the authors provide an ablation study removing the CCB while keeping the other components unchanged?
  2. How well does DART generalize to defenses outside AutoAdvExBench, particularly new defenses without existing published attacks? Could the authors provide additional evidence or analysis?
  3. Could the authors compare DART with other automated adaptive attack methods or a simpler agent workflow under similar settings?
2 comments
AuthorsOpenReview ↗

Response to Reviewer bvxv (1/2): CCB Contribution and Generalization (CW1–2 / Q1–Q2)

Reviewer Concerns Addressed — CW1–2 / Q1–Q2

CW1 / Q1 — CCB contribution: The reviewer asks whether CCB provides a distinct benefit and whether removing it or supplying the same profiling evidence without structured cues produces comparable results.

CW2 / Q2 — Generalization: The reviewer asks whether DART generalizes beyond AutoAdvExBench, especially to defenses without published bypass attacks.

Our Response

We thank Reviewer bvxv for these suggestions. We conducted additional experiments on (1) CCB utilization and representation and (2) external generalization and mechanism-level attack transfer. Full results appear in Shared Responses (1/4)–(3/4).

1. CCB Contribution (CW1 / Q1)

In Shared Response (1/4), we compare four CCB configurations across 51 defenses using Claude Opus. Removing CCB from both ART and OpenWorld decreases Success from 27/51 (52.9%) to 13/51 (25.5%) and completed evaluations from 51/51 to 34/51.

Extending CCB from ART-only to both stages increases Success from 22/51 (43.1%) to 27/51 (52.9%) and reduces mean paired robust accuracy (RA) from 44.16% to 35.60% (−8.56 pp). Both evaluate all 51 defenses, showing an attack-strength gain beyond improved completion.

To isolate representation effects, Shared Response (2/4) compares Structured CCB with Flattened Profiling Observations. Both retain identical diagnostic measurements, backbone, other prompts, retrieval tools, candidate limits, attack procedures, and budgets. Flattened removes the eight-category hierarchy while retaining field names and values.

Both achieve Success on the same 27 defenses, with similar mean RA over 48 jointly completed targets (33.05% Structured; 32.83% Flattened). Thus, the hierarchy alone does not demonstrate an additional attack-strength gain.

However, Structured CCB completes 51/51 vs. 48/51 evaluations, with lower recorded runtime (36.62 vs. 50.5 h) and API cost (USD 99.46 vs. USD 119.63). These results separate the contribution of using diagnostic evidence from the observed completion and resource benefits associated with its structured organization.

2. Generalization Beyond AutoAdvExBench (CW2 / Q2)

In Shared Response (3/4), we evaluate five external defenses spanning different datasets and mechanisms. DART completes 5/5 evaluations (zero Failures) and achieves 3/5 Successes (60%):

  • TopoReformer: APGD-DLR; RA 0.0%.
  • ZeroPur: BPDA + CW-Margin PGD; RA 9.5%.
  • LeapNet-1: EOT + Margin-PGD; RA 28.5%.

On LeapNet-1, identifying randomized execution and adapting an EOT-based attack reduces RA from 55.0% to 28.5% (−26.5 pp), changing Not Broken to Success. DART also executes adapted attacks and obtains valid RA for GradNorm/Swin-B and RMAggNet, although both remain Not Broken.

We additionally examine three in-benchmark defenses for which our literature search found no dedicated published bypass: pReLU_ICML24, adversarial_detector, and Advanced-Gradient-Obfuscating/GD. Based on observed gradient reliability, rejection behavior, and randomized preprocessing, DART achieves RA of 0.0%, 0.0%, and 1.5%, respectively.

Specifically, DART applies direct gradient optimization when gradients are reliable, avoids unnecessary detector-specific optimization when rejection is inactive, and combines BPDA with EOT for randomized, non-differentiable preprocessing.

Together, these evaluations support DART's ability to analyze new implementations and operationalize relevant attack principles. The in-benchmark cases provide evidence of mechanism-level transfer distinct from external generalization. We will clarify these findings and their scope in the revision.

Question 3 — Alternative Automated Methods

We address Q3 (CW2) separately in Response (2/2): Comparison with Automated Adaptive Attack Methods, including a new comparison with A3 using its official implementation.

AuthorsOpenReview ↗

Response to Reviewer bvxv (2/2): Comparison with Automated Adaptive Attack Methods (CW2 / Q3)

Update (18 Nov): All 35 compatible A3 evaluations are complete, replacing the previously reported interim results on 30 defenses.

Reviewer Concerns Addressed — CW2 / Q3

The reviewer requests comparisons with established automated adaptive attack methods and simpler agent workflows.

Our Response

We thank Reviewer bvxv for this suggestion. We compare DART with the AutoAdvExBench agent Baseline and A3 (Yao et al., 2021), distinguishing attack strength, evaluation coverage, runtime, and manual integration.

1. Simpler Agent Baseline

The Baseline uses fixed forward-pass, differentiable-forward, FGSM, and iterative-attack stages without defense profiling, CCB guidance, or external retrieval.

DART increases Success from 3/51 (5.9%) to 22/51 (43.1%) with Qwen and 16/51 (31.4%) to 27/51 (52.9%) with Claude Opus. DART completes 51/51 evaluations, versus 16/51 and 22/51 for Baseline, while reducing LLM tokens by 84% and 95%, respectively.

2. A3 Setup and Evaluation Coverage

Using A3's official implementation, we searched attack sequences, losses, hyperparameters, and transformations, including BPDA. Both methods used the same 200 images for final RA scoring on the original defense under benchmark-defined ℓ∞ constraints. Success requires RA < 0.5 × clean accuracy.

A3 requires differentiable PyTorch interfaces, which we manually constructed for 35/51 defenses. The remaining 16 could not be evaluated in our A3 environment: 15 TensorFlow-1 defenses require conversion, and the A5 forward script failed in our A3 setup.

All 35 A3 evaluations are now complete. On these defenses, A3 breaks 15 and DART breaks 13. Importantly, DART also completes all 16 A3-incompatible defenses, successfully attacking 14, achieving 27/51 successes with 51/51 valid evaluations without manual interface construction.

The 16 incompatible defenses are excluded from paired comparisons, not counted as A3 failures.

3. Paired Attack Strength (35 Defenses)

MethodSuccessMean RA ↓
A315/35 (42.9%)39.3%
DART (Claude Opus)13/35 (37.1%)44.8%

A3 is stronger on the matched subset, breaking all 13 defenses broken by DART plus DISCO (A3/DART RA: 2.0%/61.5%) and ATLD (25.0%/59.5%).

However, DART achieves lower RA on TWIS (51.5%/41.5%), Ensemble Everything Everywhere (66.5%/60.0%), Hedge Defense (67.5%/61.5%), and Reversal via NS (68.5%/63.0%).

For SABRE-v2 (MNIST, ε = 0.031), A3 failed to generate a non-zero perturbation, yielding 100.0% RA, whereas DART reduced RA to 82.0% (−18.0 pp). Although neither met the Success criterion, DART demonstrated stronger attack effectiveness on this defense.

4. Runtime and Manual Integration

We compare summed per-defense execution times, not parallel wall-clock time.

MetricA3DART (Opus)
Completed evaluations35/5151/51
Total runtime (h)367.736.62
Average runtime (h/defense)10.50.72
Per-defense runtime range (h)0.2–45.4—
Manually constructed attack-side interfaces350

DART achieves approximately 15× lower reported average runtime per defense, while completing all 51 evaluations. A3's runtime excludes manual interface construction and restarted attempts.

This is not a compute-matched comparison: defense sets, hardware, and search budgets differ. A3 uses 64 TPE trials, 100 search samples, and three sequential attacks with successive halving.

DART automates defense profiling, attack-side interface construction, attack selection, adaptation, and final evaluation.

5. Conclusion and Revision

A3 achieves stronger attacks on the 35 matched defenses (15 vs. 13 successes), whereas DART provides broader autonomous evaluation. DART completes all 51 heterogeneous defenses, including 16 unavailable to our A3 setup, with 14 successful attacks, without manual interface integration and with substantially lower reported average runtime.

These results reinforce DART's contribution: automatically translating defense-specific understanding into executable adaptive attacks across heterogeneous implementations, rather than uniformly outperforming specialized attack search.

We will incorporate the completed A3 comparison, paired CA/RA, compatibility limitations, budgets, runtime, and integration requirements into the revised manuscript.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Shared Response (1/4) — CCB Utilization and Attack Effectiveness

We thank reviewers bvxv, 7vzB, and WWgr for their constructive feedback. Because several concerns overlap, we organize our additional experiments into four shared responses: (1) CCB Utilization and Attack Effectiveness, (2) Structured CCB vs. Flattened Profiling Observations, (3) Generalization Beyond AutoAdvExBench, and (4) Evaluation Stability and Reproducibility. Each response identifies the reviewer concerns it addresses; reviewer-specific comments cover paired Baseline–DART outcomes, query accounting, budgets, and alternative attack methods.

Reviewer concerns addressed

  • bvxv — Critical Weakness 1 / Question 1: What does CCB contribute, and what happens if it is removed while the remaining workflow is retained?
  • 7vzB — Critical Weaknesses 1–2: How much of the improvement comes from CCB utilization, and how can attack strength be distinguished from evaluation completion? The matched representation-only test requested in Question 1 is addressed in Shared Response (2/4).

1. CCB Utilization Across ART and OpenWorld

Using Claude Opus, we evaluated four CCB configurations on the same 51 AutoAdvExBench defenses. Success requires final robust accuracy (RA) below 50% of clean accuracy; success rates use all 51 defenses as the denominator.

ConfigurationART / OpenWorld CCBSuccessCompleted RA
No CCBNo / No13/51 (25.5%)34/51 (66.7%)
ART-only CCBYes / No22/51 (43.1%)51/51 (100.0%)
OpenWorld-only CCBNo / Yes24/51 (47.1%)50/51 (98.0%)
Full DARTYes / Yes27/51 (52.9%)51/51 (100.0%)

Removing CCB from both stages reduced successful attacks from 27 to 13 and completed evaluations from 51 to 34.

2. Attack Strength vs. Evaluation Coverage

We compared paired RA only on defenses where both configurations completed evaluation:

Full DART vs.Paired DefensesComparator Mean RAFull DART Mean RAChange (pp)
No CCB3443.72%39.32%−4.40
ART-only CCB5144.16%35.60%−8.56
OpenWorld-only CCB5037.55%34.56%−2.99

Each row uses its own paired set. Lower RA indicates a stronger attack. In particular, ART-only and Full DART both completed 51/51 evaluations, yet success increased 22 → 27 and mean RA decreased 44.16% → 35.60% (−8.56 pp). Thus, the gain is not solely due to recovering failed evaluations.

Of the 17 No-CCB failures, Full DART converted 11 to Success and six to Not Broken; four further No-CCB Not Broken defenses became Success. Relative to ART-only CCB, six Not Broken defenses became Success.

These results demonstrate gains from utilizing CCB evidence in both attack stages, including improved attack strength at unchanged evaluation coverage. The separate contribution of hierarchical CCB formatting, tested with identical profiling evidence, is examined in Shared Response (2/4). Baseline–DART paired outcomes and resource budgets are addressed in our reviewer-specific responses.

AuthorsOpenReview ↗

Shared Response (2/4) — Structured CCB vs. Flattened Profiling Observations

Reviewer concerns addressed

  • bvxv — Critical Weakness 1: Would the same profiling information work equally well without explicit structured CCB categories?
  • 7vzB — Critical Weakness 1 / Question 1: Can the structured CCB be replaced by flattened profiling observations while keeping other prompts, tools, candidate limits, and execution budgets unchanged?

We thank both reviewers for requesting this controlled comparison. Unlike the CCB-utilization ablation in Shared Response (1/4), this experiment changes the representation of the same diagnostic evidence, not whether the agent receives that evidence.

1. Representation-Controlled Evaluation

We compared Structured CCB with Flattened Profiling Observations using Claude Opus on the same 51 defenses, with 200 images per defense. The backbone, remaining prompts, retrieval tools, candidate limits, attack procedure, and configured budgets were held constant.

MetricStructured CCBFlattened Observations
Attack Success (all 51)27/51 (52.9%)27/51 (52.9%)
Attack-code generation51/51 (100.0%)51/51 (100.0%)
Completed final RA51/51 (100.0%)48/51 (94.1%)
Not Broken / Failure24 / 021 / 3
Mean RA on common 4833.05%32.83%

Flattened Observations removes the hierarchy of eight CCB categories but retains the attack-relevant diagnostic field names and numerical values. For example, Structured CCB groups grad_l1=25041.97, frac_zero=0.0, and vanishing=false under probe_gradient_masking; Flattened retains these named observations without their category block.

Both conditions succeeded on exactly the same 27 defenses. Across the 48 jointly completed defenses, Structured had lower RA on 16, Flattened on 17, and 15 were tied. Hence, attack strength was comparable when the diagnostic evidence was held fixed; we do not claim a separate attack-strength gain from hierarchy alone.

2. Completion and Resource Consumption

Both conditions generated attack code on all 51 defenses. However, Flattened did not obtain final RA on MemoryDef/baseline-multik, vits-robustness-torch/baseline, and yoon_2021/baseline: each reached the 240-minute execution limit after code generation. Structured CCB completed all three and classified them as Not Broken.

MetricStructured CCBFlattened Observations
Completed final RA51/5148/51
Total recorded execution time36.62 h50.5 h
Total LLM API costUSD 99.46USD 119.63

On the 48 common completed defenses, sums of the rounded per-defense cost entries were approximately USD 93.80 (Structured) and USD 119.64 (Flattened). Structured cost less on 36 defenses, Flattened on 10, with two ties. The minor difference between rounded-entry sums and reported totals comes from rounding.

In summary, structured organization did not change which defenses were broken, but was associated with higher evaluation completion and lower runtime and API expenditure. We will distinguish evidence-utilization effects from representation effects in the revised paper.

AuthorsOpenReview ↗

Shared Response (3/4) — Generalization Beyond AutoAdvExBench

Reviewer concerns addressed

  • bvxv — Critical Weakness 2 / Question 2: Does DART generalize beyond AutoAdvExBench, including to defenses without dedicated published bypass attacks?
  • WWgr — Single-benchmark generalization concern: Can the approach work across other defense implementations and datasets?

We thank both reviewers for these questions. We evaluated five external defenses and examined three in-benchmark cases of mechanism-level attack transfer.

1. Five External Defenses

With 200 images per defense and Success defined as RA below 50% of clean accuracy, DART completed 5/5 evaluations (zero Failures) and achieved 3/5 Successes (60%).

Defense (Dataset)MechanismAdapted AttackClean / RA (%)Outcome
TopoReformer (MNIST) [5]Topological purificationAPGD-DLR98.5 / 0.0Success
ZeroPur (CIFAR-10) [6]Two-stage purificationBPDA + CW-Margin PGD94.0 / 9.5Success
LeapNet-1 (GTSRB) [7]Dynamic layer routingEOT + Margin-PGD96.5 / 28.5Success
GradNorm/Swin-B (ImageNet) [10]Gradient-norm regularizationCW/Margin PGD75.5 / 56.5Not Broken
RMAggNet (MNIST) [4]Reed–Muller codes / rejectionContinuous-score surrogate100.0 / 73.0Not Broken

Not Broken does not mean execution failure: GradNorm/Swin-B and RMAggNet both yielded valid RA from executable adapted attacks. GradNorm's outcome is consistent with the benchmark's robust Vision Transformer (vits-robustness-torch) and more than ten adversarial-training defenses that remained Not Broken.

RMAggNet was tested under L∞ ε = 0.1, stricter than ε = 0.3 generally used for MNIST in our benchmark. A continuous-score surrogate addressing coding/rejection operations reduced RA 85.0% → 73.0% (Tier 1 → final), without reaching Success.

Attack selection followed defense mechanisms: APGD-DLR for TopoReformer's differentiable path [3], BPDA for ZeroPur's purification, and EOT for LeapNet-1's randomized execution [1]. On LeapNet-1, adaptation reduced RA 55.0% → 28.5% (−26.5 pp), changing Not Broken to Success.

2. In-Benchmark Mechanism Transfer

For three defenses without a dedicated published bypass identified in our literature search, DART selected attacks based on diagnostics:

DefenseDiagnostic EvidenceAdaptationFinal RA
pReLU_ICML24 [8]No masking, stochasticity, or rejectionMargin-based white-box attack0.0%
adversarial_detector [2]Rejection inactive; usable gradientsMargin Auto-PGD0.0%
Advanced-Gradient-Obfuscating/GD [9]Randomized, non-differentiable preprocessingBPDA + EOT1.5%

These cases show DART selecting direct gradients when reliable, classifier optimization when rejection is inactive, and BPDA/EOT for randomized preprocessing. The external evaluations and transfer cases provide evidence of defense-specific attack adaptation beyond the original benchmark.

References

[1] Athalye, A., Carlini, N., & Wagner, D. (2018). Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples. ICML.

[2] Carlini, N., Debenedetti, E., Rando, J., Nasr, M., & Tramèr, F. (2025). AutoAdvExBench: Benchmarking Autonomous Exploitation of Adversarial Example Defenses. ICML, PMLR 267, 6778–6793. arXiv:2503.01811.

[3] Croce, F., & Hein, M. (2020). Reliable Evaluation of Adversarial Robustness with an Ensemble of Diverse Parameter-Free Attacks. ICML, PMLR 119, 2206–2216.

[4] Fentham, D., Parker, D., & Ryan, M. (2024). Using Reed-Muller Codes for Classification with Rejection and Recovery. FPS 2023, LNCS 14551, 36–52. arXiv:2309.06359.

[5] Kumar, B., Aravinthakashan, A. S., Satyanarayan, A., Gakhar, I., & Verma, U. (2025). TopoReformer: Mitigating Adversarial Attacks Using Topological Purification in OCR Models. arXiv:2511.15807.

[6] Liu, E., Yang, Z., Liu, B., Meng, X., Bi, X., Han, J., & Xiao, B. (2026). ZeroPur: Succinct Training-Free Adversarial Purification. IEEE Transactions on Image Processing, 35, 9547–9559.

[7] Ma, Z., Luo, X., Song, Q., & Tan, R. (2025). Dynamic Layer Routing Defense for Real-Time Embedded Vision. ACM Transactions on Embedded Computing Systems, 24(5s), Article 144.

[8] Min, H., & Vidal, R. (2024). Can Implicit Bias Imply Adversarial Robustness? ICML, PMLR 235, 35687–35718.

[9] Qiu, H., Zeng, Y., Zheng, Q., Zhang, T., Qiu, M., & Memmi, G. (2020). Mitigating Advanced Adversarial Attacks with More Advanced Gradient Obfuscation Techniques. arXiv:2005.13712.

[10] Rodríguez-Muñoz, A., Wang, T., & Torralba, A. (2024). Characterizing Model Robustness via Natural Input Gradients. ECCV.

AuthorsOpenReview ↗

Shared Response (4/4) — Evaluation Stability and Reproducibility

Reviewer concerns addressed

  • 7vzB — Critical Weakness 4 / Question 4: How were the 200 images selected, and are the main outcomes stable across independent agent runs, larger evaluation sets, and near-threshold cases?
  • WWgr — Evaluation-variance concern: Can the authors provide run-to-run variability estimates and error bars?

We thank both reviewers for these suggestions. We report four complementary checks: (1) image-selection procedure, (2) three independent agent runs, (3) 200-versus-1,000-image evaluations, and (4) an outcome near the success threshold.

1. Evaluation Image Selection

For each AutoAdvExBench dataset, we sorted the 1,000 evaluation images by filename and selected the first 200, matching labels from labels.txt. For ImageNet, supported image extensions were filtered before sorting. There was no random sampling or selection based on attack results, and comparisons on the same dataset used identical image subsets.

2. Run-to-Run Stability: Three Independent Agent Executions

Using Claude Opus, we reran the full DART workflow—defense profiling, attack generation, and final RA scoring—three times on each of five defenses, retaining the same 200 images and attack constraints.

DefenseRun 1 RARun 2 RARun 3 RAMean RA ± SD
TRADES54.5054.5054.5054.50 ± 0.00
kWTA-Activation1.507.503.004.00 ± 3.12
Mixup-Inference5.501.505.004.00 ± 2.18
adversarial_detector0.500.500.000.33 ± 0.29
DiffPure88.5089.5087.5088.50 ± 1.00

RA is in percent; sample SD is in percentage points. All 15/15 runs produced valid RA and each defense retained the same Success/Not Broken outcome across three runs. kWTA-Activation showed the largest RA range (6.0 pp), but all three runs remained Success.

3. Evaluation-Set Stability: 200 vs. 1,000 Images

We also evaluated final attacks on 1,000 instead of 200 images across all 51 defenses.

MetricResult
200-image evaluation27 Success / 24 Not Broken
1,000-image evaluation27 Success / 24 Not Broken
Outcome changes0/51
Mean absolute RA difference1.00 pp
Median absolute RA difference0.60 pp
Maximum absolute RA difference6.10 pp

TurningWeaknessIntoStrength had the largest RA change, 41.5% → 35.4%, while remaining Success.

4. Near-Threshold Outcome

For obfuscated-gradients/sap, the success criterion is RA < 0.5 × clean accuracy:

  • 200 images: Clean 87.0%, threshold 43.5%, RA 44.0% → Not Broken.
  • 1,000 images: Clean 86.2%, threshold 43.1%, RA 44.8% → Not Broken.

The original RA was only 0.5 pp above the threshold, but the outcome remained unchanged with 1,000 images.

5. Summary and Manuscript Revision

These results establish deterministic subset selection, consistent outcomes across 15 independent full-agent runs on five defenses, and zero outcome changes when increasing evaluation size across all 51 defenses, including a near-threshold example. We will add the image-selection protocol, run-level RA and sample SD, expanded-set results, and near-threshold analysis to the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.