acceptodds
Under review as a conference paper at ICLR 2027

Exploiting Video Temporal Capability as an Attack Pathway in Multimodal Large Language Models

Abstract

Multimodal Large Language Models (MLLMs) are rapidly gaining advanced video understanding capabilities, including temporal localization, cross-segment reasoning, and long-context integration. However, existing video jailbreak attacks largely treat video as a rich carrier of harmful content, leaving these emergent capabilities underexplored as an attack pathway. In this work, we introduce **Video Capability-Safety Trap (VCST)**, whose key insight is to make harmful content recovery a prerequisite for completing a legitimate video understanding task, inducing an intrinsic conflict between capability-driven task completion and safety-driven refusal. VCST temporally disperses harmful information across video segments and requires the model to localize and integrate it, reducing frame-level exposure while co-opting the model's video reasoning capabilities to advance the attack objective. Extensive experiments on a diverse set of MLLMs show that VCST consistently outperforms prior video-based attacks across both non-thinking and thinking settings, and exhibits significantly stronger robustness to existing defenses. To counter VCST, we further develop a video-oriented defense that encourages deliberation on harmful intent detection. Additional analyses validate the reliability of our evaluation and examine the effect of model scale. Our findings highlight an emerging capability-safety tension in MLLMs and call for safety alignment that explicitly accounts for evolving video capabilities.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multimodal jailbreak, Multimodal model safety, Video-based attack, Black-box jailbreak, Video understanding

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_214xOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how video temporal reasoning capabilities can be exploited to jailbreak multimodal large language models. It proposes VCST, which disperses harmful instructions across video segments and embeds their reconstruction within temporal localization and cross-segment reasoning tasks. Experiments across multiple models show higher attack success rates than existing baselines, including in thinking mode, while also examining defense robustness and video-oriented detection strategies.

Critical strengths
  • Important and timely video safety problem.
  • Effective and computationally inexpensive attack design.
  • Extensive comparisons across models and reasoning modes.
Critical weaknesses
  • The novelty of VCST is insufficiently established against closely related multimodal jailbreak methods that also distribute harmful information and require its reconstruction.

  • The central capability-safety mechanism is not convincingly isolated, and the defense evaluations do not adequately establish robustness against stronger video-aware safeguards.

  • The attack evaluation lacks sufficient statistical validation, while the absence of released code and several implementation ambiguities limit reproducibility.

  • The paper has ethical concerns regarding inconsistencies in human-subject involvement, insufficient disclosure of annotator protections, and unclear responsible disclosure procedures for affected model providers.

Questions for authors
  1. The novelty of VCST is insufficiently established against closely related multimodal jailbreak attacks. [1] already distributes harmful information across visual components and guides the model to reconstruct the underlying request, while [2] similarly decomposes harmful queries across multiple images. Moreover, TempJail (2026), which is cited but not discussed, also exploits temporal arrangements of adversarial information in videos. What is the fundamental technical distinction between VCST and these approaches beyond arranging fragmented harmful content temporally? The authors should provide a more rigorous novelty discussion and include comparisons with directly related methods, particularly HIMRD.

  2. The proposed capability-safety trap is not sufficiently isolated by the ablation experiments. Table 6 shows that removing temporal splitting reduces average ASR only from 81.7% to 78.8%, while the unsplit variant still outperforms MCV. Furthermore, the VCST prompt in Figure 19 contains linguistic framing similar to the strategies criticized in Section 3.2.1. Could the authors separately evaluate the VCST prompt with an unsplit typographic image, remove the contextual framing from the prompt, and apply the VCST prompt to MCV videos? These controls are important for determining whether the improvement comes from exploiting video temporal reasoning or from a stronger textual jailbreak prompt.

  3. The defense evaluation does not adequately establish robustness against stronger video-aware safeguards. Figure 3 relies on a random single-frame filter, which is particularly favorable to an attack that places harmful content in only a small fraction of the video. Table 7 already shows considerably higher detection rates under key-frame inspection. Could the authors evaluate multi-frame sampling, video-wide OCR, and key-frame filtering under different inspection budgets? In addition, the proposed tri-step defense achieves relatively low VCST detection rates in Table 4. Please report its actual post-defense ASR and false-positive rates on benign text-bearing videos to establish whether it provides meaningful protection.

  4. The attack success evaluation requires stronger validation. The binary safety judge includes policy categories such as legal, financial, and health advice, which do not necessarily correspond to actionable harmful content. Appendix A.1 also includes reasoning traces in safety evaluation, whereas Section 4.2 discusses improvements in final-response safety under thinking mode. Could the authors report final-response-only ASR and use a graded harmfulness metric, such as StrongREJECT [3], to distinguish actionable harmful outputs from broad policy violations? Category-level results would further clarify whether the reported improvements hold for clearly harmful behaviors.

  5. The experimental reliability and generality need stronger evidence. VCST randomly selects natural video contexts and insertion positions, but the main results do not report variation across independent attack constructions or random seeds. Some ASR improvements over competing methods are also relatively small. Could the authors provide confidence intervals and statistical significance tests across multiple runs? The closed-source results should also be interpreted more carefully, since DMN outperforms VCST on Gemini-3.1-Pro and Claude-Sonnet-5 frequently produces generic restart responses. Evaluation on a video-native safety benchmark such as Video-SafetyBench would further strengthen the generality of the conclusions.

  6. The paper does not provide the implementation code, substantially limiting reproducibility. VCST involves several implementation-sensitive components, including harmful-query decomposition, typographic rendering, temporal splitting, random clip insertion, prompt construction, and safety evaluation. Without the code, it is difficult to verify whether the baselines follow comparable protocols or reproduce the reported attack success rates. The paper should also clarify the exact splitting orientation and position, which are described inconsistently across Figure 4, Figure 19, and Appendix A.2.

  7. Minor: Several presentation and bibliography issues require correction. Figure 3 uses inconsistent y-axis ranges, Figure 15 contains the typo "FLM-4.6V-Flash", and some appendix figure references are incorrect. The bibliography also contains inaccurate publication information, such as the venue detail for VRSA (should be CVPR finding). There misses a space before page range.

[1] Ma, Teng, et al. "Heuristic-Induced Multimodal Risk Distribution Jailbreak Attack for Multimodal Large Language Models." ICCV. 2025.

[2] Yang, Zuopeng, et al. "Distraction is All You Need for Multimodal Large Language Model Jailbreaking." CVPR. 2025.

[3] Souly, Alexandra, et al. "A StrongREJECT for Empty Jailbreaks." NeurIPS. 2024.

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

This paper proposes VCST, a video-based jailbreak attack that exploits temporal reasoning to bypass safety alignment, achieving high attack success rates and strong robustness against existing defenses.

Critical strengths
  1. The paper is good written.

  2. The experiments are overall solid.

Critical weaknesses
  1. The current experiments mainly demonstrate that the individual design choices improve downstream attack performance. However, I do not think they directly establish that these improvements stem from the models' video temporal understanding capabilities. This is my primary concern, and addressing it may require a substantial amount of additional experimental work.

  2. SPTV is also worth including as a baseline for video-based attacks.

  3. The defense experiments only consider MCV as the primary attack baseline. Including more attack baselines would make the results more convincing.

Questions for authors

Please see the Weaknesses section for details. I would consider raising my score if my concerns are adequately addressed, particularly the first one.

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

This paper investigates the security vulnerabilities of multimodal large language models (MLLMs) arising from their increasingly sophisticated video understanding capabilities. The authors propose Video Capability-Safety Trap (VCST), a jailbreak framework that exploits temporal localization and cross-segment reasoning to induce harmful responses. Specifically, VCST decomposes a harmful instruction into complementary visual segments, inserts them at different temporal positions within natural video contexts, and formulates a video understanding task that requires the target model to locate and integrate the distributed information before generating its response. The authors argue that this process creates a conflict between capability-driven task completion and safety-driven refusal. Experiments on multiple open-source and proprietary MLLMs demonstrate that VCST achieves higher attack success rates than existing image- and video-based jailbreak methods under both thinking and non-thinking settings.

Critical strengths
  1. The paper explores an interesting research question: whether the capabilities acquired by MLLMs for legitimate video understanding tasks can themselves become attack surfaces. Rather than simply using videos to carry malicious instructions, the proposed framework attempts to exploit temporal localization and cross-segment information integration as part of the jailbreak process. This capability-oriented perspective is potentially valuable for understanding emerging security risks as multimodal models become more capable.

  2. The experimental results demonstrate that VCST is an effective jailbreak method across multiple MLLM architectures. According to Table 2, VCST achieves average ASRs of 79.9% and 76.3% in non-thinking and thinking modes, respectively, outperforming the strongest competing baselines by 8.5 and 15.4 percentage points. The relatively small performance degradation under thinking mode is particularly interesting, suggesting that additional reasoning does not necessarily eliminate vulnerabilities induced by task-oriented multimodal inputs. The additional evaluations on proprietary models further support the practical relevance of the attack.

Critical weaknesses
  1. The central claim that video temporal capabilities cause the jailbreak vulnerability is not sufficiently established. The main novelty of the paper lies in attributing the success of VCST to the exploitation of video temporal localization and cross-segment reasoning. However, the proposed implementation primarily decomposes a typographic harmful instruction into two complementary static clips and inserts them into an otherwise unrelated video sequence. This construction does not necessarily require sophisticated video understanding, since the same information can potentially be recovered through multi-image comparison or visual text reconstruction.

  2. The proposed attack does not clearly separate legitimate video understanding from explicit harmful-content generation. The paper argues that VCST makes recovering harmful information a prerequisite for completing a legitimate video understanding task, thereby creating a conflict between capability and safety objectives. However, the actual attack prompt in Appendix Figure 19 goes considerably beyond temporal localization and content reconstruction. It explicitly instructs the model to infer missing behavioral steps, provide detailed explanations, and generate concrete actions using real-world knowledge when the video lacks sufficient information. Consequently, the final task is not merely to understand or reconstruct video content; it explicitly encourages the model to generate detailed procedural information that is absent from the original video. This distinction matters because the observed jailbreak success may be partially attributable to conventional instruction-following manipulation rather than an intrinsic conflict between video understanding and safety alignment.

Questions for authors

see the weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.