acceptodds
Withdrawn from ICLR 2027

Where and How to Steer Representations for Jailbreak Attacks on Multimodal Large Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Jailbreak attacks can induce Multimodal Large Language Models (MLLMs) to respond to harmful requests. Although numerous jailbreak attacks have been proposed, the majority of existing methods rely on output-oriented objectives to elicit affirmative responses. Such objectives are superficial and coarse-grained, which overlook the more fundamental internal representations associated with the model's refusal behavior. Meanwhile, some high-performing attacks jointly perturb visual and textual inputs, increasing the complexity of adversarial optimization. To address these limitations, we aim to identify an internal state-oriented optimization objective and develop an effective image-only attack, with a particular focus on determining where and in which direction adversarial optimization should be performed. First, to determine the attack position, we systematically analyze the representations at different positions of MLLMs. Second, to seek a suitable optimization direction, we find and validate a refusal-associated direction that can effectively influence the model's behavior. Building on these two insights, we propose a jailbreak attack that optimizes bounded perturbations to steer the representation at the target position away from the refusal-associated direction. Experiments on three MLLMs show that our method achieves attack success rates of %, %, and % on Llama-3.2-11B, Qwen2.5-VL-7B, and MiniGPT-4-13B, respectively. Despite perturbing only the input image, our method outperforms the baselines under the joint image-and-text attack settings. These results highlight the effectiveness of leveraging internal representations to guide targeted jailbreak attacks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multimodal Large Language Models, Multimodal Safety, Internal Representations

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_cqcwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an image-only white-box jailbreak for MLLMs. Using matched harmful/harmless pairs, it analyzes four representation positions and picks the post-instruction token. It then selects a refusal direction (rejected minus accepted mean) through activation steering, and finally optimizes an ℓ∞ = 8/255 image perturbation that pushes the representation away from this direction. ASR is 95.65/93.04/98.69 on Llama-3.2-11B, Qwen2.5-VL-7B and MiniGPT-4-13B, slightly above JailBound with joint image+text perturbation.

Critical strengths
  1. The position analysis with matched pairs is clean. MiniGPT-4 shows no behaviour-aligned pattern and also refuses much later (median first refusal token 133.5 vs. 21/3), which nicely links representation to behaviour.
  2. The attack is simple, and the direction still works on held-out categories (Table 2).
Critical weaknesses
  1. ASR seems to be keyword-based. App. A.2 says the refusal templates are used "to evaluate attack success", while §6.4 mentions JailBound's judge template. If it is keywords, off-topic or misread outputs count as success; No-Attack ASR is already 73% on Llama. The refusal direction is also built from the same keyword labels used for scoring. In typo images the harmful phrase is inside the image, so the perturbation may simply stop the model from reading it. A judge-based harmfulness metric is needed.
  2. The claimed contribution is the objective (internal state vs. output likelihood), but no output-likelihood image attack is run under the same ε, steps and evaluator. The gap between JailBound (I) at 54.81 and 93.04 on Qwen is too large to accept without such a control. The ablation shows L_perp and L_reg add less than 1 point, so the method is close to existing refusal-direction attacks [1, 2], and these should be discussed.

[1] Refusal in Language Models Is Mediated by a Single Direction. NIPS, 2024. [2] Uncovering Safety Risks of Large Language Models through Concept Activation Vector. NIPS, 2024.

Questions for authors
  1. The "where" and "how" are chosen from similarity analysis and activation addition, but the attack itself is never run with the alternatives. What ASR do you get when the same attack targets the question-last token or the image tokens, or uses the input-based direction?
  2. Is the matched multimodal pair construction necessary? Does a refusal direction computed from text-only harmful/harmless prompts, as in [1], give similar ASR?
  3. Since the attack moves the post-instruction state along a known direction, a defender could simply monitor the projection onto d_ref. How easily are your adversarial images detected this way, and how much ASR is lost if the attack must stay below such a monitor?
  4. Does the perturbation transfer across models, or work as a universal perturbation?

[1] Refusal in Language Models Is Mediated by a Single Direction. NIPS, 2024.

Reviewer_796nOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the intent–action gap in multimodal large language models, where models can recover harmful intentions hidden across images, text, and task frames but still comply with unsafe requests. The authors introduce BIND-5K, a dataset and benchmark comprising matched harmful and benign examples to evaluate intent-aware safety decisions. They further propose BIND-Align, which combines intent-first supervised fine-tuning with reward-based reinforcement learning. Experiments across four MLLM backbones demonstrate improved jailbreak resistance while largely preserving helpfulness on benign requests.

Critical strengths

1. An interesting perspective on multimodal safety failures. The paper highlights the intent–action gap in MLLMs, showing that models may comply with harmful requests even after recovering their underlying intent. The analysis provides useful insights into multimodal safety failures.

2. A meaningful benchmark design with promising results. BIND-5K uses counterfactually matched examples to jointly evaluate safety and helpfulness, discouraging blanket refusal. Experiments across multiple backbones demonstrate improved jailbreak resistance while largely preserving benign helpfulness.

Critical weaknesses

W1. The case for targeting internal representations is asserted, not shown. The paper calls output-oriented objectives "superficial and coarse-grained" (abstract; lines 078-080), but never demonstrates this. There is no experiment in which an output-oriented loss falls short and the representation loss does better. The second argument, that joint image-text optimization is complex and expensive (lines 081-083), is an argument for image-only attacks. It says nothing about which loss to use, because an output-oriented loss can also be optimized on the image alone. The efficiency benefit is also never measured. As written, I can't tell what concrete problem the representation objective solves.

W2. The "where" and "how" choices are never tested as attacks. For "where": Section 4 finds that both the question-last and post-instruction positions track refusal behavior, then picks post-instruction mainly because it sits closer to generation (lines 274-277). The two positions are never compared as attack targets. For "how": Section 5 shows that adding the response-based direction raises refusal more than adding the input-based one (Figure 5). That is evidence about inducing refusal. The attack does the opposite: it pushes the representation away from the direction through an image perturbation. A direction that is better for the first job need not be better for the second. Since the paper never runs the attack with each position and each direction, its two central choices rest on proxy analyses.

Questions for authors

Q1. Could you compare your objective with an output-oriented one (e.g., maximizing the likelihood of an affirmative prefix) in the same image-only setting, with the same perturbation budget and number of steps? This is the most direct way to support the motivation in lines 078-080.

Q2. Could you report ASR when the attack targets the question-last position instead of post-instruction, and when it uses the input-based direction instead of the response-based one? A small 2x2 table on one model would be enough.

Q3. Eq. (7) applies the same α to d_input and d_output, which are raw mean differences (Eqs. 5-6) and may have quite different norms. If d_output simply has a larger norm, Figure 5 would favor it for that reason alone. Are the directions normalized? If not, does the result still hold once they are normalized or the intervention norm is matched?

Q4. Appendix A.4 notes that MiniGPT-4-13B does not show the behavior-related pattern at the post-instruction position, yet it reaches the highest ASR (98.69%). If the attack works just as well without that pattern, what does the representation analysis contribute to choosing the target? On this model, the choice seems to rest only on the position being the last token before generation.

Q5. The gains over JailBound (I+T) are 1.27, 1.64, and 1.50 points. Could you report the mean and standard deviation over several random seeds so readers can judge whether these gaps are meaningful? The introduction also argues that joint image-text attacks are costly; per-sample runtime and GPU memory compared with JailBound would back this up.

Reviewer_MF6sOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a representation-guided jailbreak attack against multimodal large language models. Instead of directly optimizing the model's output, the method identifies an internal representation associated with refusal behavior and uses image perturbations to steer the representation away from that direction. Experiments on three models show high attack success rates. However, the novelty, evaluation reliability, and practical effectiveness need further validation.

Critical strengths
  1. Interesting perspective. The paper studies internal representations rather than relying only on output-based optimization, offering a useful perspective on multimodal safety vulnerabilities.

  2. Clear methodology. The authors systematically analyze different representation positions and compare candidate steering directions before designing the attack.

  3. Strong reported results. The method achieves high attack success rates on all three evaluated models, including 95.65% on Llama-3.2-11B and 93.04% on Qwen2.5-VL-7B.

Critical weaknesses
  1. Limited novelty. The method builds on existing work on refusal-related directions and activation steering. The main contribution is applying these ideas to image-only jailbreak optimization. The paper should better explain what fundamental insight distinguishes this approach from prior representation-based attacks.

  2. The evaluation metric may be unreliable. Attack success is judged using predefined refusal templates rather than directly checking whether the response provides harmful information. A response may contain harmful content before adding a refusal or disclaimer. The appendix reports that MiniGPT-4-13B often produces refusal expressions late in its responses, making this concern particularly relevant.

  3. The improvement over strong baselines is modest. Compared with JailBound's joint image-and-text attack, the proposed method improves overall attack success rates by only 1.27–1.64 percentage points. Some individual categories also perform worse than the baseline. Multiple runs and statistical significance tests are needed to establish that these gains are reliable.

  4. The representation analysis is not fully convincing. MiniGPT-4-13B does not show the same clear relationship between pre-generation representations and refusal behavior as the other models, yet the proposed method still achieves strong results. This raises questions about whether the identified refusal-related direction consistently explains the attack's effectiveness across different architectures.

Questions for authors
  1. Can the authors use a semantic evaluation method to verify whether the generated responses actually comply with harmful requests, rather than relying only on refusal templates?

  2. Can the authors provide results across multiple random seeds and statistical tests to demonstrate that the improvements over JailBound are significant?

  3. Why does the proposed representation-based explanation remain effective on MiniGPT-4-13B despite its weaker behavior-related representation patterns?

  4. Can the authors report per-image optimization time, perturbation visibility, and robustness under common image transformations?

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

Related papers

Discussion (0)

Sign in to comment.