acceptodds
Under review as a conference paper at ICLR 2027

DarkLLM: Learning Language-Driven Adversarial Attacks with Large Language Models

Abstract

While vision and multimodal foundation models underpin critical tasks from perception to complex reasoning, they remain highly vulnerable to adversarial attacks. However, traditional adversarial attacks are typically limited to single, predefined objectives, tightly coupling each attack to a specific model or task, which restricts their scalability and flexibility in real-world scenarios. In this work, we present DarkLLM, a novel attack framework that trains an LLM to translate natural-language attack instructions into latent attack vectors, which are then decoded into visual adversarial perturbations. By leveraging natural-language instruction tuning, DarkLLM not only unifies targeted, untargeted, segmentation, and multi-model attacks within a single framework, but also achieves flexible and controllable adversarial generation, enabling each instruction to produce a perturbation that induces desired behaviors across heterogeneous models. Through extensive experiments across 4 tasks, 13 datasets, and 15 models, we demonstrate that DarkLLM with only 1B parameters can follow attacker instructions and generate highly effective attacks against CLIP, SAM, and frontier LLMs, revealing a systemic vulnerability in modern foundation models.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
adversarial attacks, vision-language models, large language models, segment anything models

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

The paper presents DarkLLM, a language-driven framework for generating visual adversarial perturbations. An LLM interprets attack instructions and produces latent representations that condition dedicated noise generators. Through instruction tuning, surrogate-model optimization, and a decoupled training strategy, the framework supports targeted, untargeted, segmentation, and multi-model attacks. The evaluation covers multiple vision and multimodal tasks, including transfer to commercial MLLMs.

Critical strengths
  1. Unifying heterogeneous attack settings within a common framework is an important contribution. Although the individual components are largely established, their integration is well motivated and practically relevant.

  2. The experiments cover classification, retrieval, captioning, and segmentation. The segmentation results demonstrate strong transferability, while the targeted attack results show improvements over AnyAttack on several evaluated MLLMs.

  3. The appendix strengthens the empirical evaluation through analyses of unseen instruction templates, alternative target descriptions, performance on 1,000 images, threshold sensitivity, repeated API evaluations, robustness to defenses, and perceptual quality.

Critical weaknesses
  1. The perturbation generator itself appears to offer limited architectural novelty.

  2. The paper presents DarkLLM as an efficient and scalable alternative to repeated optimization, but its extensive evaluation of attack effectiveness is not accompanied by a systematic analysis of computational cost.

  3. The ablation with a frozen text encoder provides useful evidence for the effectiveness of LLM conditioning. However, since the LLM is fine-tuned while the baseline encoder remains frozen, the comparison does not fully distinguish the benefits of the LLM from those of task-specific adaptation.

  4. Table 20 appropriately evaluates two semantic-similarity thresholds. Although DarkLLM outperforms AnyAttack at both thresholds, its average ASR decreases from 55.5% at 0.3 to 11.1% at 0.5. The discussion should therefore distinguish between inducing broadly target-related content and fulfilling the complete target description.

Questions for authors
  1. What are the total training GPU-hours and latency for generating a perturbation from a new instruction? How do these compare with representative baselines, and how does reuse across images affect the amortized cost?

  2. Can a smaller or trainable text encoder provide a better effectiveness-efficiency trade-off under matched training conditions?

6 comments
AuthorsOpenReview ↗

Response to Reviewer PdWK (Part 1/2)

We sincerely thank the Reviewer PdWK for recognizing the practical relevance of our unified framework and extensive evaluations. We address your concerns below.

Q1: The perturbation generator itself appears to offer limited architectural novelty.

A1: We acknowledge the reviewer's question regarding the generator architecture. We intentionally adopted established, lightweight components for the perturbation generator to prioritize training stability and computational efficiency.

We believe that our primary novelty lies at the framework level rather than in low-level neural architectures. By introducing the Embedding as Attack mechanism, DarkLLM successfully leverages an instruction-following LLM as a central semantic controller for adversarial generation, effectively bridging natural language instructions with pixel-level adversarial noise.

In our extensive experiments, we demonstrate that even with this simple generator, DarkLLM achieves state-of-the-art or highly competitive adversarial performance across heterogeneous tasks. We consider exploring more advanced generator architectures, such as Diffusion Transformers (DiT), as a promising direction to further enhance its generative capacity.

Q2: The discussion should distinguish between inducing broadly target-related content and fulfilling the complete target description.

A2: We sincerely appreciate the reviewer's careful reading and the constructive suggestion. We agree that distinguishing between these semantic levels is crucial. The scoring criteria in Appendix Table 11 provide a concrete basis for this distinction:

  • Threshold 0.3 (Broad target-related content): Our criteria define scores between 0.3 and 0.5 as sharing the "same general theme but different descriptions". Thus, [email protected] directly measures the attack's ability to induce the model to generate broadly target-related content.
  • Threshold 0.5 (Subject-level semantic alignment): Scores between 0.5 and 0.7 strictly require the "same subject". Therefore, [email protected] enforces a much stronger subject-level constraint, ensuring that the core semantics of the generated text align closely with the exact target description.

The decrease in average ASR from 55.5% to 11.1% explicitly demonstrates that enforcing strict subject-level alignment is substantially more challenging than inducing broad semantic redirection. Nevertheless, DarkLLM consistently outperforms AnyAttack across both difficulty levels.

Following your advice, we will incorporate this criteria-based discussion into Section C.2 to explicitly clarify the differences between broad semantic redirection and strict subject-level alignment.

Q3: What are the total training GPU-hours and latency for generating a perturbation from a new instruction? How do these compare with representative baselines, and how does reuse across images affect the amortized cost?

A3: We appreciate the opportunity to clarify the computational requirements. Developing the unified DarkLLM framework requires an upfront, one-time training of approximately 1,040 GPU-hours (130 hours on 8 H200 GPUs). Once trained, generating a perturbation for a new instruction requires only a single forward pass, taking just 0.25 seconds on a single H200 GPU.

To contextualize this, it is essential to compare the time complexities between different adversarial paradigms. Sample-wise targeted attacks (e.g., M-Attack) scale with O(X×Y)\mathcal{O}(X \times Y) complexity, as they require iterative backpropagation for every combination of XX images and YY targets. For instance, optimizing a perturbation for a single image-target pair takes roughly 40 seconds using M-Attack on a single H200 GPU.

In contrast, the computational advantage of DarkLLM emerges from perturbation reuse. Because DarkLLM generates an image-agnostic Universal Adversarial Perturbation (UAP), its generation cost scales with O(Y)\mathcal{O}(Y) during inference. The 0.25-second generation cost is incurred only once per target instruction, and applying this perturbation across XX images requires no further optimization.

To illustrate the practical efficiency, consider a standard robustness evaluation setup using X=1,000X=1,000 images and just Y=100Y=100 target instructions. A sample-wise approach would require optimizing 100,000 independent pairs, resulting in approximately 1,111 GPU-hours of computation. For DarkLLM, the total computational cost is 1,040 GPU-hours for training plus merely 25 seconds for inference. Therefore, XX and YY do not need to be significantly large. DarkLLM becomes computationally advantageous even at moderate evaluation scales.

AuthorsOpenReview ↗

Response to Reviewer PdWK (Part 2/2)

Q4: Can a smaller or trainable text encoder provide a better effectiveness-efficiency trade-off under matched training conditions?

A4: We sincerely appreciate this insightful suggestion. To investigate this, we conducted the requested experiment by unfreezing the CLIP ViT-L/14 text encoder and training it under matched conditions. The results are summarized below:

Conditioning ModelGPT-4o ASR (%)GPT-5-mini ASR (%)Gemini-2.5-Flash ASR (%)Average
Frozen Text Encoder62.050.059.057.0
Trainable Text Encoder63.655.664.661.3
DarkLLM67.758.663.663.3

As the results indicate, fine-tuning the text encoder yields improvements over the frozen baseline, raising the average ASR from 57.0% to 61.3%. However, DarkLLM maintains a clear advantage in overall attack effectiveness. Although the trainable encoder performs competitively on Gemini-2.5-Flash, DarkLLM consistently outperforms this baseline on both GPT-4o and GPT-5-mini by clear margins, achieving a higher overall average ASR.

A trainable text encoder indeed offers a more computationally efficient alternative. However, relying solely on conventional text encoders fundamentally limits the framework's capacity for deep semantic understanding and flexible interaction, which is our core motivation for exploring an LLM-driven adversarial framework.

Indeed, our framework is motivated by the broader success of foundational LLMs. Today, LLMs have become the cornerstone of numerous successful applications, enabling breakthroughs in areas such as complex visual reasoning (e.g., LLaVA, NeurIPS ’23), image generation (e.g., BAGEL, arXiv ’25; OmniGen2, CVPR ’26), and robotic manipulation (e.g., OpenVLA, CoRL ’24; RT-2, CoRL ’23).

In the adversarial domain, previous works have explored using CLIP (e.g., AnyAttack, CVPR ’25; CGNC, ECCV ’24) or diffusion models (e.g., DiffAttack, TPAMI ’24) for adversarial generation. DarkLLM takes a fundamentally different path. To the best of our knowledge, DarkLLM is the first work in the field to leverage an LLM for generating visual adversarial perturbations.

Crucially, empowering an LLM to perform diverse adversarial attacks is non-trivial. To make this possible, we conducted careful architectural designs and systematic optimizations for each specific attack objective to maximize performance. Furthermore, this architecture provides many possibilities for future exploration. Promising directions include integrating reinforcement learning, incorporating Diffusion Transformers (DiT), and enabling joint outputs of visual perturbations and adversarial text. We believe our framework and empirical results will provide the community with valuable insights.

AuthorsOpenReview ↗

Reviewer-Author Discussion Period Ends in ONE Day

Dear Reviewer PdWK,

Thank you again for your valuable review. We have provided a detailed response to address the concerns you raised. As the discussion period deadline is approaching, we would be very grateful if you could revisit our clarifications. Please let us know if any questions or ambiguities remain. We are eager to engage in further discussion and would be more than happy to address any additional questions.

We sincerely appreciate your time and effort.

Warm regards,

Authors of Submission 9607

Reviewer_eH7ROpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies an adversarial attack method that unifies untargeted attack, targeted attack, and SAM. The author used an LLM as an instruction-based text encoder, along with a global noise embedding. Then the author used separate noise generators for different scenarios. Experiments show comparable results compared to the single-task attack methods.

Critical strengths
  1. This paper provides a sufficient ablation study of different surrogate models and parameters.

  2. The paper provides extensive experiments across different domains, tasks, and adversarial attack baselines.

  3. The setting of the paper is quite substantial. It tries to explore a unified framework for adversarial attacks across different tasks, instructions, and models.

Critical weaknesses

Overall, although this paper claims that the language-driven aspect is substantial, experiments show that the main effect actually comes from the consecutive UAP generator and the training recipe.

  1. On the task axis, this paper is not scalable and is not flexible since each task has its own generator, and adding a new task requires retraining the generator.

  2. The flexible adaptation to different attack instructions seems to apply only to targeted attacks. The untargeted attack and the attack on the SAM model actually do not differ much from a UAP, since these instructions, as shown in Fig 2 and 5, do not have parameters, i.e., they are constant. Therefore, the hidden state from the large language model appears to be a constant as well. In this sense, at least in these two tasks, the optimization target of DarkLLM is the reparameterization of the traditional UAP.

    • In Tab. 1, the average is 76.5, not distinguishable from X-Transfer base
    • In Tab. 22, removing the LLM as the embedding generator yielded comparable, even better performance
  3. On Targeted attack, DarkLLM is more efficient compared to sample-wise attack, while more effective compared to universal adversarial attack methods. However, to what extent is DarkLLM more efficient, (e.g, when sample-wise adversarial attack needs to adapt XX images and Y targets) is not illustrated. As far as I am concerned, this method consumes substantial computational resources. Therefore, XX and YY need to be significantly large.

Questions for authors

Overall, are there any further analyses that show other advantages of using an LLM as the shared embedder, such as

  1. generalization to out-of-distribution targets (labels that are outside COCO), a more diverse prompt set (the prompt set in the appendix seems very close to the original ones)

  2. positive transfer by adding more task domains, e.g., adding SAM attack to targeted attack can improve both performance

From what the paper presents, it seems like a simpler engineering solution that uses a text classifier, which routes to different attack methods (X-Transfer in Tab. 1, Standalone Generator in Tab. 22, M-attack in Tab. 2), can also yield comparable performance while also providing a unified interface?

5 comments
AuthorsOpenReview ↗

Response to Reviewer eH7R (Part 1/3)

We sincerely thank the Reviewer eH7R for recognizing the value of our unified framework and extensive evaluations. Below, we clarify our contributions and address your concerns.

Q1: On the task axis, this paper is not scalable and is not flexible since each task has its own generator, and adding a new task requires retraining the generator.

A1: We appreciate this question. In fact, this decoupled design is precisely the mechanism that enables scalability in our framework.

In adversarial learning, optimizing for fundamentally different objectives, such as pixel-level segmentation boundaries versus global multimodal semantics, introduces gradient interference. Forcing a single generator to handle all tasks leads to catastrophic forgetting and degraded performance.

By assigning a lightweight, task-specific generator (which adds negligible parameter overhead) while sharing the LLM controller, DarkLLM achieves Plug-and-Play Modularity. Adding a new task does not require retraining the entire framework or destroying previously learned capabilities. Instead, it simply involves introducing a new lightweight generator branch and applying parameter-efficient fine-tuning (LoRA) to the LLM, making it scalable to new tasks.

Q2: The flexible adaptation to different attack instructions seems to apply only to targeted attacks. In Tab. 22, removing the LLM as the embedding generator yielded comparable, even better performance.

A2: We appreciate this observation. While specialized single-task generators excel in their isolated domains, the ultimate goal of DarkLLM is to demonstrate the potential of LLMs to unify diverse adversarial attacks within a single framework. To this end, we selected representative targeted and untargeted attacks against the most widely used foundation models today (CLIP, SAM, and MLLMs).

Regarding the results in Table 22, we analyze the necessity of the LLM from two perspectives:

  • Semantic Conditioning: While a standalone generator achieves comparable performance on the SAM attack, it is fundamentally incapable of processing conversational instructions or handling targeted attacks. These targeted attacks require the model to parse and embed diverse, open-ended target semantics. The LLM serves as the crucial semantic controller, bridging heterogeneous attacks via a single natural language interface.
  • Negative Transfer: In multi-task adversarial training, combining fundamentally different objectives typically leads to negative transfer. Table 22 actually validates the effectiveness of our decoupled architecture: it demonstrates that integrating the SAM attack into a multi-task LLM framework preserves its performance, without degradation compared to a specialized, single-task model.

Q3: To what extent is DarkLLM more efficient, (e.g, when sample-wise adversarial attack needs to adapt X images and Y targets) is not illustrated.

A3: We appreciate this feedback. To quantify the efficiency trade-off, it is essential to compare the O(X×Y)\mathcal{O}(X \times Y) complexity of sample-wise attacks with DarkLLM’s O(Y)\mathcal{O}(Y) inference cost.

Sample-wise attacks require iterative optimization for every instance. Based on our tests (e.g., using M-Attack), generating a perturbation takes roughly 40 seconds per image-target pair on an H200 GPU. In contrast, DarkLLM incurs an upfront training cost of approximately 1,040 GPU-hours (130 hours on 8 GPUs). However, generating a UAP for a new target requires only a single forward pass taking 0.25 seconds. Crucially, applying this image-agnostic UAP to XX images requires no additional optimization.

To illustrate the practical efficiency, consider a standard robustness evaluation setup using X=1,000X=1,000 images and Y=100Y=100 target instructions. A sample-wise approach would require optimizing 100,000 independent pairs, resulting in approximately 1,111 GPU-hours of computation. For DarkLLM, the total computational cost is 1,040 GPU-hours for training plus merely 25 seconds for inference. Therefore, XX and YY do not need to be significantly large. DarkLLM becomes computationally advantageous even at moderate evaluation scales.

AuthorsOpenReview ↗

Response to Reviewer eH7R (Part 2/3)

Q4: Generalization to out-of-distribution targets (labels that are outside COCO), a more diverse prompt set (the prompt set in the appendix seems very close to the original ones).

A4: We appreciate the reviewer's careful observation. We address the generalization to open-ended targets and the evaluation on a more diverse prompt set respectively below:

  • Generalization to Out-of-Distribution Targets: We would like to kindly note that the generalization experiments to open-ended targets have been included in Appendix C.2, where two complementary experiments demonstrate DarkLLM's capability beyond standard COCO annotations:

    • Unseen target descriptions (Tables 16–17): We used Qwen3-VL 8B to generate new, context-rich captions for the target images, which differ significantly from the standard COCO style and were completely unseen during training. When conditioned on these alternative descriptions without retraining, DarkLLM maintained ASRs of 57.0%–68.0% across five commercial MLLMs, confirming its effectiveness on OOD targets.
    • Expanded evaluation (Table 20): To validate our framework's effectiveness across a wider variety of target concepts, we scaled the evaluation to 1,000 random source-target image pairs. Under this broader setting, DarkLLM achieved an average ASR of 55.5% at the 0.3 semantic-similarity threshold and 11.1% at the stricter 0.5 threshold, outperforming the previous SOTA method (AnyAttack) by 24.8% and 6.0%, respectively.
  • Robustness to a More Diverse Prompt Set: As noted in Tables 18–19, we evaluated five previously unseen instruction templates. The average ASR decreased by at most 1.9 percentage points, indicating that the framework successfully captures the semantic meaning of the objective rather than memorizing fixed syntactic patterns.

    To further address your concern, we designed a diverse set of prompts featuring drastically different syntaxes (e.g., direct instructions, goal specifications, and question forms). Evaluated without any retraining, the results are summarized below:

    A More Diverse Prompt SetGPT-4o ASR (%)GPT-5-mini ASR (%)Gemini-2.5-Flash ASR (%)
    Original Training Prompt: "Make the vision-language model identify this image as [Target]."67.758.663.6
    Diverse Prompt 1 (Direct Instruction): "Apply adversarial noise to this image so that the vision-language model describes it as [Target]."64.656.662.6
    Diverse Prompt 2 (Goal Specification): "Desired model description: [Target]. Generate an image perturbation that induces this description."66.756.662.6
    Diverse Prompt 3 (Question Form): "What perturbation would make a vision-language model describe this image as [Target] rather than describe its actual content?"65.754.566.7
    Average of the Diverse Prompt Set65.755.964.0

    Even when evaluated on this drastically distinct prompt set, DarkLLM maintains stable attack performance across all three commercial MLLMs.

AuthorsOpenReview ↗

Response to Reviewer eH7R (Part 3/3)

Q5: Positive transfer by adding more task domains, e.g., adding SAM attack to targeted attack can improve both performance.

A5: We appreciate the reviewer's perspective on positive transfer, which is highly desirable in standard multi-task learning. However, in the context of unifying heterogeneous adversarial attacks, the fundamental challenge lies in highly conflicting optimization objectives.

Specifically, a targeted MLLM attack optimizes for a global semantic shift to steer language generation, whereas a SAM attack disrupts fine-grained, pixel-level boundary features. Because these two objectives target fundamentally different feature spaces, our early joint-training experiments revealed that forcing them to share a single generator typically leads to gradient interference and negative transfer, degrading performance on both ends rather than improving it.

Therefore, the primary technical achievement of DarkLLM is not achieving positive transfer between conflicting domains, but rather successfully mitigating negative transfer. By employing a decoupled architecture and a progressive training strategy, DarkLLM allows these highly divergent attacks to coexist seamlessly within a single LLM controller, a crucial step toward scalable, unified adversarial frameworks.

Q6: A simpler engineering solution that uses a text classifier, which routes to different attack methods can also yield comparable performance while also providing a unified interface?

A6: We appreciate this alternative design. While a "classifier + router" is a viable engineering solution, relying solely on text classifiers fundamentally limits the framework's capacity for deep semantic understanding and flexible interaction, which is our core motivation for exploring an LLM-driven adversarial framework.

Technically, a simple classifier-routing architecture faces two critical bottlenecks:

  • Semantic Loss: A classifier outputs only a discrete label, discarding the rich, open-ended target semantics required for generative attacks. DarkLLM avoids this by utilizing the LLM's hidden states to seamlessly encode dense semantics directly into the perturbation.
  • Inference Inefficiency: Routing to existing targeted methods (e.g., M-Attack) requires sample-wise iterative optimization for every new image. In contrast, DarkLLM learns a direct mapping that generates targeted UAPs via a single forward pass in 0.25 seconds.

We would like to further address your concerns regarding our contributions by elaborating on the motivation behind our approach. Our framework is motivated by the broader success of foundational LLMs. LLMs have become the cornerstone of numerous successful applications today, enabling breakthroughs in areas such as complex visual reasoning (e.g., LLaVA, NeurIPS ’23), image generation (e.g., BAGEL, arXiv ’25; OmniGen2, CVPR ’26), and robotic manipulation (e.g., OpenVLA, CoRL ’24; RT-2, CoRL ’23).

In the adversarial domain, previous works have explored using CLIP (e.g., AnyAttack, CVPR ’25; CGNC, ECCV ’24) or diffusion models (e.g., DiffAttack, TPAMI ’24) for adversarial generation. DarkLLM takes a fundamentally different path. To the best of our knowledge, DarkLLM is the first work in the field to leverage an LLM for generating visual adversarial perturbations.

Crucially, empowering an LLM to perform diverse adversarial attacks is non-trivial. To make this possible, we conducted careful architectural designs and systematic optimizations for each specific attack objective to maximize performance. Furthermore, this architecture provides many possibilities for future exploration. Promising directions include integrating reinforcement learning, incorporating Diffusion Transformers (DiT), and enabling joint outputs of visual perturbations and adversarial text. We believe our framework and empirical results will provide the community with valuable insights.

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

The paper trains a 1B model called DarkLLM to map a templated instruction to a controlled token. The token's hidden state conditions a generator an output Universal Adversarial Perturbation (UAP) for targeted/untargeted, or SAM/ mutti-modal attack.

Critical strengths
  1. The idea of using natural language as a common interface for the multiple attack scenarios are practical. Single UAP can attack attack and ten can transferred to VLM like CLIP and MLLM, which traditional attack method cannot.

  2. The experimental results are good. Several design choices are ablated: BCE vs MSE for cross-backbone SAM transfer, multi-prompt ensembling, decoupled progressive training, and surrogate scaling.

Critical weaknesses
  1. The method selects one of the 4 control token. The generator ignores the image and attends over a single conditioning token (the paper admits the attention weights are identically one). Each token routes to its own dedicated generator. The LLM is effectively choosing among four classes, plus embedding a caption for the targeted one. The paper never tests whether it “decides” or “reasons” about anything. There are no ambiguous, compositional, or out-of-set instructions. The claim of using LLM's semantic reasoning is over claim in terms of evidence presented.

  2. Missing baselines and comparison with the closed architecture. Closest text-conditioned generators (closest architecture of DarkLLM), a stronger universal targeted baseline, and prior “language as attack interface” work are missed. CGNC (ECCV 2024) (listed in Introdcution but compared). C-PGC (ICCV 2025) (One Perturbation is Enough: On Generating Universal Adversarial Perturbations against Vision-Language Pre-training Models), UnivIntruder (“One Surrogate to Fool Them All”)

  3. Flexibility claim. The authors concede transfer to new tasks is not established. Target captions are drawn from COCO in training and evaluation, so open-ended targets are untested.

Questions for authors
  1. What happens with compositional or never-seen target types?
  2. How does the method compare to a text-encoder-conditioned generator trained identically?
  3. How does DarkLLM compare to a CGNC-style generator with a plain text encoder trained identically?
  4. What is new relative to Instruct2Attack, InSUR and Con Instruction beyond using an LLM as the conditioning encoder?
4 comments
AuthorsOpenReview ↗

Response to Reviewer ZHKp (Part 1/2)

We sincerely thank the Reviewer ZHKp for recognizing our core contributions and extensive evaluations. We address your concerns below.

Q1: The claim of using LLM's semantic reasoning is over claim in terms of evidence presented.

A1: We appreciate the reviewer's feedback. We agree that "semantic reasoning" typically implies step-by-step logical deduction, which may overstate the LLM's role in our pipeline. We will replace this term with "semantic understanding and conditioning" in the revised manuscript to more accurately reflect its function.

We would also like to elaborate on the dual role of the LLM in our architecture, which extends beyond a simple 4-way task classifier. On one hand, the generated control token (e.g., [VLM_TGT]) indeed serves as a routing mechanism to select the appropriate generator branch. More importantly, the contextualized hidden states of the LLM extract and carry dense semantic information from the natural language instruction. This is particularly evident in the targeted MLLM attacks. As illustrated in Figure 8 and Figure 14, completely different target descriptions produce distinct latent representations within the same generator branch, thereby conditioning entirely different pixel-level perturbations. Therefore, the LLM contributes essential semantic conditioning that translates open-ended target intents into specific adversarial signals, going far beyond simple task classification.

Q2: Missing baselines and comparison with the closed architecture.

A2: We thank the reviewer for pointing out these missing baselines. We will cite and include a comparison of these works in the revised version.

  • Comparison with C-PGC: We respectfully note that C-PGC was already included as a key baseline in Table 1 (Section 4.2) of our original manuscript for untargeted VLM attacks. As demonstrated, DarkLLM outperforms C-PGC across all metrics. For instance, on zero-shot classification, DarkLLM achieves an average ASR of 76.5% compared to C-PGC's 64.0%. Similarly, DarkLLM shows clear advantages in image-text retrieval and captioning.

  • Comparison with CGNC and UnivIntruder: Following your suggestion, we evaluated CGNC and UnivIntruder on the targeted image captioning task against commercial MLLMs. The results are summarized below:

    MethodGPT-4o (ASR %)GPT-5-mini (ASR %)Inference Time per Target
    CGNC1.01.0~5.8ms
    UnivIntruder68.754.5~81 mins (5000 iters)
    DarkLLM (Ours)67.758.6~0.25s (1 forward pass)

    As shown, CGNC yields an ASR of only 1.0% on the targeted image captioning task, as it is primarily designed for classification and struggles to transfer to generative multimodal tasks. Meanwhile, UnivIntruder achieves competitive effectiveness, outperforming DarkLLM by 1.0% on GPT-4o, whereas DarkLLM achieves a 4.1% higher ASR on GPT-5-mini. However, UnivIntruder relies on sample-wise iterative optimization, and requires approximately 81 minutes (5,000 steps) on a single H200 GPU to generate a single targeted perturbation. In contrast, DarkLLM generates a target-specific UAP through only a single forward pass, taking 0.25 seconds on the same GPU.

Q3: Target captions are drawn from COCO in training and evaluation, so open-ended targets are untested.

A3: We appreciate this insightful question. We would like to kindly note that the generalization experiments to open-ended targets have been included in Appendix C.2, where three complementary experiments demonstrate DarkLLM's capability beyond standard COCO annotations:

  • Unseen target descriptions (Tables 16–17): We used Qwen3-VL 8B to generate new, context-rich captions for the target images, which differ significantly from the standard COCO style. When conditioned on these alternative descriptions without retraining, DarkLLM maintained ASRs of 57.0%–68.0% across five commercial MLLMs, confirming its effectiveness on open-ended targets.

  • Unseen instruction templates (Tables 18–19): We evaluated five previously unseen instruction templates, each differing significantly from the one used during training. The average ASR decreased by at most 1.9 percentage points, indicating that the framework captures the semantic meaning of the objective rather than memorizing fixed syntactic patterns.

  • Expanded evaluation (Table 20): To validate our framework's effectiveness across a wider variety of target concepts, we scaled the evaluation to 1,000 random source-target image pairs. Under this broader setting, DarkLLM achieved an average ASR of 55.5% at the 0.3 semantic-similarity threshold and 11.1% at the stricter 0.5 threshold, outperforming the previous SOTA method (AnyAttack) by 24.8% and 6.0%, respectively.

AuthorsOpenReview ↗

Response to Reviewer ZHKp (Part 2/2)

Q4: Flexibility claim. The authors concede transfer to new tasks is not established. What happens with compositional or never-seen target types?

A4: We appreciate this insightful question. To address this, it is important to distinguish between transferring to new attack tasks and handling never-seen target types within a supported task.

First, as acknowledged, transferring to a fundamentally new attack task requires a new surrogate objective and additional training. DarkLLM provides a unified language-controlled architecture, but it does not enable zero-shot transfer to entirely unobserved task domains without supervision.

However, regarding previously unseen target types within the trained task, DarkLLM exhibits strong generalization. As highlighted in our response to Q3 (and Appendix Tables 16–19), DarkLLM successfully accommodates newly generated, context-rich target descriptions. This confirms its ability to extract the underlying semantic intents beyond the fixed instruction templates or simple reference captions used during training. We will discuss this in the revised manuscript.

Q5: How does the method compare to a text-encoder-conditioned generator (or a CGNC-style generator) trained identically?

A5: We appreciate this insightful question. For this comparison, we respectfully direct the reviewer's attention to Section 4.3 of the main text. Specifically, to assess the contribution of the LLM, we replaced it with a CLIP ViT-L/14 text encoder and retrained the framework under identical settings. As shown in Figure 5(h), DarkLLM consistently outperforms this text-encoder baseline across all evaluated commercial MLLMs, increasing the average ASR from 57.0% to 63.3%.

Q6: What is new relative to Instruct2Attack, InSUR and Con Instruction beyond using an LLM as the conditioning encoder?

A6: We thank the reviewer for highlighting these related works. We will cite and discuss them in the revised manuscript.

  • Instruct2Attack performs language-guided semantic attacks through an adversarially guided diffusion process conditioned on both a source image and an editing instruction. DarkLLM instead learns to generate a norm-bounded additive perturbation without receiving the source image, allowing the exact same perturbation to be reused across different images.
  • InSUR improves diffusion-based semantic adversarial generation by addressing instruction uncertainty, sampling stability, and constraints in 2D/3D scenarios. DarkLLM focuses on an entirely different formulation: directly decoding instruction-dependent representations into reusable UAPs for heterogeneous visual attack objectives.
  • Con Instruction optimizes adversarial images or audio to align with textual instructions in the embedding space for multimodal jailbreaking. DarkLLM, however, learns an instruction-to-perturbation mapping during training, meaning inference for a new target description does not require any instruction-specific iterative optimization. Furthermore, its evaluated objectives are much broader, unifying targeted/untargeted vision-language attacks, segmentation attacks, and joint multi-model attacks.

In summary, the novelty of DarkLLM extends beyond merely replacing the conditioning encoder. We would like to further address your concerns regarding our contributions by elaborating on the motivation behind our approach. Our framework is motivated by the broader success of foundational LLMs. LLMs have become the cornerstone of numerous successful applications today, enabling breakthroughs in areas such as complex visual reasoning (e.g., LLaVA, NeurIPS ’23), image generation (e.g., BAGEL, arXiv ’25; OmniGen2, CVPR ’26), and robotic manipulation (e.g., OpenVLA, CoRL ’24; RT-2, CoRL ’23).

In the adversarial domain, previous works have explored using CLIP (e.g., AnyAttack, CVPR ’25; CGNC, ECCV ’24) or diffusion models (e.g., DiffAttack, TPAMI ’24) for adversarial generation. DarkLLM takes a fundamentally different path. To the best of our knowledge, DarkLLM is the first work in the field to leverage an LLM for generating visual adversarial perturbations.

Crucially, empowering an LLM to perform diverse adversarial attacks is non-trivial. To make this possible, we conducted careful architectural designs and systematic optimizations for each specific attack objective to maximize performance. Furthermore, this architecture provides many possibilities for future exploration. Promising directions include integrating reinforcement learning, incorporating Diffusion Transformers (DiT), and enabling joint outputs of visual perturbations and adversarial text. We believe our framework and empirical results will provide the community with valuable insights.

AuthorsOpenReview ↗

Reviewer-Author Discussion Period Ends in ONE Day

Dear Reviewer ZHKp,

Thank you again for your valuable review. We have provided a detailed response to address the concerns you raised. As the discussion period deadline is approaching, we would be very grateful if you could revisit our clarifications. Please let us know if any questions or ambiguities remain. We are eager to engage in further discussion and would be more than happy to address any additional questions.

We sincerely appreciate your time and effort.

Warm regards,

Authors of Submission 9607

AuthorsOpenReview ↗

Reviewer-Author Discussion Period Ends in 8 HOURS

Dear Reviewer ZHKp,

The reviewer-author discussion period is going to end in less than 8 hours. We are eager to hear your feedback and comments on our responses. We are eager to address any remaining questions.

We would also like to take a moment to share the motivation behind DarkLLM.

Empowering a single LLM to unify completely different attacks, from disrupting pixel-level segmentation to steering multimodal semantics, was indeed non-trivial. During our early experiments, we struggled with gradient conflicts that caused interference between different attack objectives.

We did not give up on this because we genuinely believe the community needs a language-driven framework to simplify safety evaluations, much like how LLMs have unified tasks in general AI. Our goal was not to chase absolute performance against specialized, single-task attacks, but to explore whether a foundational LLM could serve as a central "semantic engine" to coordinate highly heterogeneous adversarial intents.

While we acknowledge that DarkLLM is an early step and does not resolve all theoretical challenges, the strong empirical results gave us a reason to be hopeful. We observed that by adopting a decoupled architecture, DarkLLM can indeed translate open-ended natural language into distinct, effective pixel-level perturbations without catastrophic interference. We sincerely believe our explorations can bring valuable insights and a new perspective to the community.

We sincerely appreciate your time and effort.

Warm regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.