acceptodds
Under review as a conference paper at ICLR 2027

MoMART: Mixture of Multimodal Auto-Red-Teamers

Abstract

Multimodal large language models (MLLMs) are trained for safety across the input modalities they support, yet whether these per-modality defenses remain effective under coordinated multimodal attacks remains poorly understood. We present **MoMART** (Mixture of Multimodal Auto-Red-Teamers), an adaptive compositional red-teaming framework that decomposes harmful behaviors across text, image, audio, and video using reusable modality specialists built from atomic transformations. MoMART searches over admissible modality subsets, composes the resulting artifacts into a single query, and iteratively refines attacks through a rating-guided feedback loop. Evaluating on the HarmBench benchmark against six frontier models, we find that multimodal composition consistently outperforms the strongest constituent modality under matched query budgets. On Gemini-Omni, attack success increases from for the best single modality to when all four modalities are composed, with ASR increasing monotonically across composition steps. Robustness to composition varies substantially across model families, with Claude and GPT exhibiting stronger resistance than Qwen and Llama. Beyond aggregate ASR, behavior-level matching reveals 29 and 40 joint-only successes on Gemini-Omni and Qwen-Omni, respectively, that are not achieved by any constituent modality alone. These results show that independent modality evaluation can miss vulnerabilities that emerge only through joint use, motivating adaptive, composition-aware safety assessment.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multimodal jailbreaking, Automated red-teaming, Compositional attacks

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_hW53OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript proposes Mixture of Multimodal Auto-Red-Teamers (MoMART), an automated black-box red-teaming framework for evaluating modality subsets under matched conditions. MoMART combines modality-specific jailbreak strategies across text, image, audio, and video, and iteratively refines attacks using feedback from an LLM-based judge. Experiments on HarmBench demonstrate higher attack success rates than existing baselines, with additional analyses showing that multimodal composition can expose vulnerabilities not observed in individual modalities.

Critical strengths

The manuscript provides systematic evaluations of modality combinations and behavior-level analyses that offer useful insights into vulnerabilities arising from jointly delivered multimodal attacks. However, concerns remain regarding the methodological novelty and the fairness of experimental comparisons.

Critical weaknesses

1. Limited Methodological Novelty and Missing Baselines

MoMART primarily combines existing modality-specific jailbreak techniques with feedback-guided refinement, without introducing a substantially new attack mechanism or optimization algorithm. Although the manuscript discusses related compositional methods such as HACA and JMLLM, these methods are not included in the experimental comparisons. Comparisons with such approaches under compatible modality settings are important for establishing MoMART’s contributions beyond existing multimodal jailbreak frameworks.

2. Incomplete Attack-Budget Control and Efficiency Evaluation

The shared retry limit applies to each behavior-strategy trial rather than the overall attack procedure, and auxiliary LLM calls are not budget-matched across methods. Consequently, the substantial ASR improvements in Table 2 are difficult to assess without comparisons under matched aggregate attack budgets. Furthermore, Table 9 reports target queries to first success only for successful behaviors, without measuring end-to-end runtime or the computational costs of multimodal artifact generation. These limitations make it difficult to determine whether MoMART’s performance and efficiency advantages hold under comparable overall computational budgets.

Questions for authors

Please see the weaknesses above.

Reviewer_dMJ1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces MoMART, a black-box red-teaming framework that combines modality-specific attack transformations across text, image, audio, and video. It enumerates admissible modality subsets and refines generated inputs through a judge-guided feedback loop using a fixed strategy pool. Evaluations on 200 HarmBench behaviors across six targets report higher attack success than the evaluated baselines. On two omni-modal targets, four-modality attacks outperform the strongest single modality by 14.5 and 20.5 percentage points, with behavior-level analysis identifying 29 and 40 joint-only successes. Additional experiments examine a larger SafeBench behavior set, attacker and judge substitutions, input-guarded targets, and generative outputs. These results suggest that single-modality evaluation can miss failures exposed by joint inputs, although the controls do not fully isolate the contribution of multimodal composition.

Critical strengths

The systematic modality-subset evaluation is the strongest contribution. Comparing a joint attack against its strongest constituent modality is more informative than comparing only against text attacks. The behavior-level joint-only analysis further examines whether composition reaches failures outside the observed single-modality success sets. This provides a useful direction for evaluating omni-modal systems. The framework is evaluated through a realistic black-box interface and shows substantial attack effectiveness across the tested targets. Attacker and judge substitutions provide some robustness checks, while adaptive attacks against the combined input-guard and downstream-model system strengthen the practical relevance of the results. These experiments support the usefulness of the framework as an evaluation tool, even though its methodological differentiation remains limited.

Critical weaknesses

The methodological novelty is limited, and the baseline comparison does not sufficiently establish the framework’s incremental contribution. MoMART combines established feedback-guided search, modality-specific attacks, and atomic strategy composition within a fixed library. The paper explicitly builds on related compositional approaches, yet closely related multimodal methods such as HACA, JMLLM, and Arondight are absent from the main comparison. Outperforming text-search and single-modality methods therefore does not adequately demonstrate an advance over existing multimodal composition frameworks. The all-subset evaluation is useful, but the paper needs stronger comparative evidence to establish a sufficiently distinct contribution. The experiments provide insufficient controlled validation of the claim that the gains arise specifically from multimodal composition. Matching the retry limit per strategy does not match total search effort, particularly when modality subsets contain different numbers of strategies. Controls for repeated same-modality content, comparable input information, and a shared total query budget are missing. The paper also lacks controlled ablations of feedback, specialist coordination, and simple composition of existing attacks; observational action rankings do not establish component contributions. The larger-scale and generative-output studies demonstrate reuse but do not resolve these attribution issues. Consequently, the reported effectiveness is more convincing than the explanation of why the framework improves performance.

Questions for authors
  • Table 6 reports different strategy counts across subsets, whereas Table 9 specifies 15 strategies per subset. Which strategy counts are used for Tables 2–4? Does the query-efficiency metric include queries spent on unsuccessful strategies before the first success?

  • Can composition be compared against repeated same-modality inputs and simple concatenation of independently generated attacks under a shared total query budget and comparable information content? These controls would help determine whether cross-modal coordination provides additional value beyond repetition and broader search.

  • Figure 2 reports two-modality joint-only rates of 12.0% and 12.9%. Averaging the six two-modality joint-only counts in Tables 13–14 over 200 behaviors instead gives 22.0% and approximately 23.6%. What aggregation or additional filtering explains this discrepancy?

Reviewer_Tq8wOpenReview ↗
Rating
2(−1)
Confidence
3(−1)
Informativeness
2(−1)
Presentation
2
Soundness
2
Read review
Paper summary

MoMART is a black-box multimodal red team testing framework that combines text, image, audio, and video attack modules, continuously adjusting attacks through a unified scoring feedback mechanism. It enumerates the modality combinations supported by the target model from a fixed strategy library, keeping the selected combination unchanged during each attack.

The paper tested 200 HarmBench harmful behaviors on six target models and evaluated all 15 non-empty modality combinations on two full-modal models. The experiments compared joint attacks with their strongest single-modality attacks, reporting four-modal successful cases that 29 and 40 single-modality attacks failed to cover. In addition, the paper also tested different attackers and evaluators, larger-scale benchmarks, generative outputs, and models with input defenses.

Critical strengths
  1. Systematic Evaluation of Modal Combinations The paper tests all 15 non-empty modal combinations on two full-modality models and compares each combination with the strongest single-modality attack within that combination, rather than only with text attacks. Behavior-level analysis further differentiates between cases that succeed collectively and those unique to combined attacks. Four-modality attacks found 29 and 40 successful cases not covered by single-modality experiments, respectively. This suggests that evaluating only a single modality may miss security risks present under combined inputs.

  2. Comprehensive Experimental Coverage The paper conducts tests on 6 models using HarmBench, supplemented with alternative attackers, HarmBench classifiers, and larger-scale SafeBench experiments. Appendix M also evaluates input defenses alongside downstream models, rather than merely filtering existing attacks. These experiments enhance the practical value of the research but do not prove that the conclusions generalize to all models.

Critical weaknesses

The experiment did not sufficiently distinguish the effects of modal combinations, nor did it fully explain query efficiency.

First, the limit of 20 retries applies to each action-strategy combination, while different modal combinations in Table 6 contain different numbers of strategies. However, both audio single-modality and four-modality attacks include 15 strategies, so the difference in the number of strategies cannot directly explain the performance gap between them. It is still necessary to clarify the complete strategy pool, the 15-strategy setting in Table 9, and the relationship with the total number of queries actually consumed by each action. Table 9 only records the number of queries for successful actions, which cannot fully reflect the search cost.

Second, each selected modality receives the full harmful action, but the experiment did not control for the total input content or the number of repetitions.

Therefore, the results demonstrate that joint input is more effective under the current experimental setup, but they do not fully distinguish the effects of cross-modal interactions from those of additional attack content. This mainly limits the interpretation of the source of performance and efficiency, without denying that joint input may pose extra security risks.

Questions for authors
  1. Number of strategies and query costs: How are the results of each strategy aggregated into the ASR of the modal combination? Which experiments use the full number of strategies in Table 6, and which use the 15-strategy setting in Table 9? What is the total number of queries for each action, and does it include failed attempts?
  2. Comparison of related methods: Is it possible, under the same text-image interface, action set, evaluator, and total query budget, to compare with HACA or other similar combination attack methods? Can the input content volume or repetition degree be controlled to distinguish the effects of modal combination from additional attack content?
  3. Single-modality control approach: Section 4.3.1 states that completely identical and unmodified constituent materials are used, while Appendix G mentions reusing single-modality attack results. In this case, is the same material from joint attacks tested separately for each modality, or are single-modality attacks independently generated and optimized? This distinction is particularly important when the attack material involves cross-modal conditioning.
Reviewer_MtezOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper examines whether multimodal LLMs trained for safety in individual modalities remain robust to attacks combining several modalities. It introduces MoMART, a black-box red-teaming framework that uses modality specialists to render harmful requests as text, images, audio and video, tests all admissible modality combinations, and refines attacks through judge feedback under a fixed retry limit. Each combined attack is compared with its strongest single-modality counterpart using the same behaviours, attacker, judge and retry limit. The experimental findings suggest that evaluating modalities separately can underestimate vulnerability, although the study only partly controls for the alternative explanations it identifies. Combined attacks still repeat the full harmful request across channels, auxiliary LLM calls are not budget-matched, success is assessed by a single LLM judge, and the ablation is limited to two omni-modal models.

Critical strengths
  1. The research question is clear, and the evaluation protocol addresses a key confound. The paper compares combined attacks with their strongest single-modality counterpart using the same behaviours, attacker, judge and retry limit, testing all 15 modality subsets on two omni-modal models (Table 3).

  2. The evaluation is broad and reveals meaningful differences between models. It covers six targets with different input interfaces, seven baselines, a second benchmark with over 2000 SafeBench behaviours, and preliminary tests on generative targets.

Critical weaknesses
  1. Methodological novelty is limited. The main contribution is an evaluation protocol rather than a new attack. MoMART builds on CoP and HACA, and its modality specialists resemble existing techniques, such as FigStep’s typographic rendering and VoiceJailbreak’s spoken narratives. The appendix also acknowledges that the study combines established transformations. Moreover, each specialist renders the full harmful behaviour in its channel, and the outputs are presented together (Section 3.3). This does not support the abstract’s claim of “decomposing” behaviours across modalities: the method duplicates rather than distributes the request, leaving genuine cross-modal interaction unestablished.

  2. Success is measured by the same LLM judge that drives the optimisation, with no external validation. The judge is both the optimisation signal and the final metric. GPT-4.1-mini provides the same rating drives the retry. An attack optimised against its own evaluator can learn to satisfy the judge, for instance with partial or fictionalised compliance, without producing genuinely harmful output.

  3. Reproducibility is limited. The paper provides neither a reproducibility statement nor a clear plan to release the code, The responsible-use appendix calls for balancing reproducibility against misuse risk but does not specify what will be shared, even under controlled access. Moreover, four of the six targets, the attacker and the judge rely on hosted APIs that may change over time. Although prices are quoted “at the time of the experiments,” the paper reports neither model snapshots nor experiment dates.

Questions for authors

I would be interested in seeing the rebuttal of the weaknesses.

Comments on the paper

1 comment

Thank you for the detailed response and additional experiments. After reconsidering the paper, I have decided to lower my rating from 3 (Weak Accept) to 2 (Weak Reject) for two reasons.

1. Incomplete comparison under a matched query budget The authors report query costs for other attack methods under a 300-query budget but do not provide their corresponding attack success rates (ASRs). In addition, Table 6 reports different strategy counts across modality combinations, while the efficiency evaluation in Table 9 uses 15 strategies. Without the ASRs from the new runs and clearer information about the evaluation settings, it remains unclear whether MoMART's original performance advantage holds under the same total query budget.

2. Limitations of the joint-only analysis Section 4.3.1 states that the comparison uses the "exact unmodified constituent artifacts." However, the authors clarify in their response that singleton attacks were generated and optimized independently. Therefore, the joint-only results show that multimodal attacks found successful cases missed by the evaluated singleton attacks. They do not directly establish that the same attack materials fail individually but succeed when combined.

Overall Assessment I acknowledge the value of the systematic multimodal evaluation and the additional experiments. However, these two limitations weaken the evidence supporting the paper's main conclusions. I therefore believe that 2 (Weak Reject) is a more appropriate rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.