acceptodds
Under review as a conference paper at ICLR 2027

Styling Past the Guard: Content-Preserving Style Attacks on VideoLLM Pipelines

10 views

Abstract

Visual style can alter the safety behavior of vision-language models without necessarily changing the content they recognize. In guarded Video Large Language Model (VideoLLM) pipelines, however, a successful attack must bypass an independent input content guard while preserving the harmful event that originally triggered detection. We present CTASO, a constrained temporal adversarial style optimization framework that searches a continuous space of video transformations under explicit content-preservation constraints. CTASO combines temporally coordinated editing, source-based content restoration, and feasibility-first optimization using input-guard feedback without querying downstream VideoLLMs. We evaluate CTASO across three content guards and four VideoLLMs, comparing it with five baselines and five adapted prior attacks. Evaluation is restricted to videos initially blocked by each guard, and event preservation is independently verified through human audit. CTASO consistently outperforms budget-matched random style search, achieving a human-audited Guard Evasion Rate of 0.61 versus 0.22 on LlavaGuard. Its effectiveness also extends to downstream harmful generation after excluding transformations that fail human audit. Comparisons with prior attacks reveal that effectiveness against unguarded models does not necessarily translate into success against a guarded pipeline. Moreover, apparent guard evasion can result from transformations that compromise the original harmful event. These findings establish content-preserving video style transfer as a distinct attack surface and highlight the importance of jointly evaluating guard evasion, event preservation, and downstream safety alignment.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
video-language model security, multimodal jailbreak, video style transfer

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_9WaYOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces CTASO, a black-box framework for generating content-preserving adversarial style transformations against guarded VideoLLM pipelines. CTASO constructs a continuous style space, applies temporally coordinated video editing, restores salient regions and text from the source video, and uses feasibility-first CMA-ES optimization to search for transformations that bypass an input guard. Experiments cover three content guards and four downstream VideoLLMs, with human auditing used to verify event preservation.

Critical strengths
  1. The paper identifies an important attack surface by distinguishing input-guard evasion from downstream VideoLLM jailbreak success.
  2. The evaluation is relatively comprehensive, covering multiple guards and VideoLLMs, with human auditing and component ablations to verify event preservation and method effectiveness.
Critical weaknesses
  1. The computational efficiency of CTASO is not adequately evaluated. The paper reports the number of candidate evaluations before the first successful attack, but this is not equivalent to time complexity or practical runtime. Each CTASO candidate requires diffusion-based video editing, saliency restoration, OCR processing, multiple similarity measurements, and possibly a guard query. Moreover, the reported mean candidate count is conditioned on successful runs, making direct efficiency comparisons potentially misleading. The paper does not report wall-clock time, GPU hours, memory consumption, or runtime scaling with video length and resolution, nor does it compare these costs with prior attacks. Such measurements are important for evaluating applicability in realistic settings.
  2. The black-box threat model assumes access to unusually informative guard feedback. CTASO requires both the guard decision and a continuous unsafe-confidence score derived from next-token logits. In many realistic systems, an external attacker receives only a binary allow/block decision and cannot access logits or calibrated confidence values. Thus, the proposed setting is stronger than a conventional decision-only black-box threat model, and it is unclear whether CMA-ES remains effective when only binary feedback is available.
  3. The preservation constraints do not directly model temporal event semantics. The automatic constraints rely primarily on frame-level CLIP similarity, salient-region correspondence, Sobel structure, and OCR consistency. These measurements can preserve appearance while failing to preserve action order, object interactions, causal relationships, or other temporally defined aspects of an event. Human audit reduces evaluation errors but is applied only after optimization and therefore cannot correct a misaligned search objective. This concern is amplified by the evaluation protocol in which the guards and VideoLLMs observe only four or eight sampled frames, potentially underrepresenting longer or transient video events.
Questions for authors
  1. Can the authors report wall-clock runtime, GPU usage, peak memory, and scaling with video duration and resolution for CTASO and the main baselines? How many diffusion-editing operations and guard queries are required when failed runs are also included?
  2. How does CTASO perform when the attacker receives only a binary allow/block response? Can the method be adapted to a strictly decision-only black-box setting without unsafe-confidence or logit access?
  3. Can the authors incorporate or evaluate temporally semantic preservation criteria, such as action recognition or video-level event representations? How sensitive are the results to the number and positions of sampled frames, particularly for long videos or short transient events?
Reviewer_G26KOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies style-based attacks against VideoLLM pipelines that use a separate input content guard. The authors propose CTASO, which changes the visual style of harmful videos while trying to preserve the original harmful content. CTASO uses temporal editing, content preservation constraints, and iterative optimization based on guard feedback. The authors evaluate CTASO on three content guards and four VideoLLMs, comparing it with several baseline and existing attack methods.

Critical strengths
  • The paper is well organized and easy to follow. I think the motivation is clear, and the overall design of CTASO is reasonable.

  • I think evaluating guard evasion and harmful event preservation together is a strength of this paper. The human evaluation is particularly useful, as it helps show that the attacks do not simply succeed by removing the harmful content from the original videos.

Critical weaknesses
  • One concern is the limited evaluation of recent models and existing VideoLLM attacks. The paper only evaluates open-source VideoLLMs and does not include API-based frontier models. I understand that API models can be costly, but testing more recent and larger open-source models would be helpful. Also, although the paper discusses recent VideoLLM attacks such as [1] and [2], it does not directly compare against them. I think including these comparisons would make the experimental evaluation stronger.

  • Another concern is the limited comparison with existing attacks, especially in terms of efficiency. CTASO requires repeated video editing and optimization, but several recent VideoLLM attacks are not included in the experiments. Although CTASO achieves high attack success rates, the current comparisons are not enough to show whether these improvements are worth the additional cost. I think comparing both attack success rates and computational costs with these methods would make the evaluation more convincing.

[1] VideoJail: Exploiting Video-Modality Vulnerabilities for Jailbreak Attacks on Multimodal Large Language Models, ICLR Workshop 2025.

[2] Breaking Multimodal LLM Safety via Video-Driven Prompting, CVPR 2026.

Questions for authors
  • How well do attacks generated by CTASO transfer across different VideoLLMs or input guards?

  • Could the authors compare CTASO with some recent VideoLLM attacks? If possible, a comparison of computational overhead would also be helpful.

  • Could the authors also evaluate CTASO against video guardrail models such as SafeWatch [1]?

[1] SafeWatch: An Efficient Safety-Policy Following Video Guardrail Model with Transparent Explanations, ICLR 2025.

Reviewer_9pHeOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CTASO, a black-box adversarial style optimization framework designed to bypass input safety guards in VideoLLM pipelines while preserving the original harmful video content. Experiments across three safety guards and four VideoLLMs demonstrate improved guard evasion and end-to-end attack success compared with existing baselines.

Critical strengths
  1. Strong empirical performance. The proposed CTASO demonstrates substantial improvements in guard evasion across three input safety guards and achieves effective end-to-end attacks against four downstream VideoLLMs. The human evaluation further supports the effectiveness of its content-preservation mechanism.
  2. Practical significance. The paper investigates an important security vulnerability in guarded VideoLLM pipelines, distinguishing input guard evasion from downstream harmful generation.
Critical weaknesses
  1. Limited evaluation against dedicated video safety guardrails. Although the paper focuses on attacking guarded VideoLLM pipelines, the evaluated input guards primarily rely on frame-level safety classification or general-purpose VLMs prompted for video safety assessment. The paper does not evaluate dedicated video safety guardrails, such as SafeWatch [1], which are specifically designed to capture temporal information and safety-relevant events.
  2. Limited applicability under decision-only black-box settings. The proposed CTASO assumes access to the input guard's unsafe-confidence scores, which provide continuous feedback for optimization. However, real-world safety systems may expose only binary block/allow decisions rather than confidence scores. The paper does not evaluate whether CTASO remains effective under such decision-only feedback, potentially limiting its applicability to practical guarded VideoLLM systems.
  3. Insufficient analysis of computational efficiency. CTASO relies on diffusion-based video editing and iterative black-box optimization, which may introduce substantial computational overhead. Although the paper reports the number of candidate evaluations required for successful attacks, it does not provide wall-clock attack time or GPU resource consumption.

[1] SafeWatch: An Efficient Safety-Policy Following Video Guardrail Model with Transparent Explanations. ICLR 2025.

Questions for authors
  1. According to weakness 1, could the authors evaluate CTASO against dedicated video safety guardrails, such as SafeWatch [1], that explicitly model temporal information and safety-relevant events? This would help establish whether the proposed attack generalizes to more advanced video safety detection systems.
  2. According to weakness 2, could the authors evaluate CTASO under a decision-only black-box setting, where the attacker receives only binary block/allow feedback rather than unsafe-confidence scores?
  3. According to weakness 3, could the authors report the average wall-clock attack time and GPU resource consumption of CTASO, preferably across different video lengths?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.