跳到正文
arXiv cs.CL· Ziqi Miao, Yi Ding, Chen Qian, Yongyao Wang, Zaoying Huang, Yuan Xiong, Ken Deng, Lijun Li, Jing Shao·· 9 小时前AI 评分38

VisCo 视觉上下文攻击:用图像驱动上下文注入越狱 MLLM

Visual Contextual Attack: Jailbreaking MLLMs with Image-Driven Context Injection

AI 导读

研究者提出 VisCo(Visual Contextual)攻击,通过四种以视觉为核心的策略构造对话上下文,并按需动态生成辅助图像,对黑盒 MLLM 实施"以视觉为中心的越狱"。

正文 · 原文

View PDF HTML (experimental)

Abstract:With the emergence of strong vision language capabilities, multimodal large language models (MLLMs) have demonstrated tremendous potential for real-world applications. However, the security vulnerabilities exhibited by the visual modality pose significant challenges to deploying such models in open-world environments. Recent studies have successfully induced harmful responses from target MLLMs by encoding harmful textual semantics directly into visual inputs. However, in these approaches, the visual modality primarily serves as a trigger for unsafe behavior, often exhibiting semantic ambiguity and lacking grounding in realistic scenarios. In this work, we define a novel setting: vision-centric jailbreak, where visual information serves as a necessary component in constructing a complete and realistic jailbreak context. Building on this setting, we propose the VisCo (Visual Contextual) Attack. VisCo fabricates contextual dialogue using four distinct vision-focused strategies, dynamically generating auxiliary images when necessary to construct a vision-centric jailbreak scenario. To maximize attack effectiveness, it incorporates automatic toxicity obfuscation and semantic refinement to produce a final attack prompt that reliably triggers harmful responses from the target black-box MLLMs. Specifically, VisCo achieves a toxicity score of 4.78 and an Attack Success Rate (ASR) of 85% on MM-SafetyBench against GPT-4o, significantly outperforming the baseline, which achieves a toxicity score of 2.48 and an ASR of 22.2%. Code: this https URL.
Subjects: Computer Vision and Pattern Recognition (cs.CV); Computation and Language (cs.CL); Cryptography and Security (cs.CR)
Cite as: arXiv:2507.02844 [cs.CV]
  (or arXiv:2507.02844v3 [cs.CV] for this version)
  https://doi.org/10.48550/arXiv.2507.02844

arXiv-issued DOI via DataCite

Journal reference: EMNLP 2025

Submission history

From: Miao Ziqi [view email]
[v1] Thu, 3 Jul 2025 17:53:12 UTC (3,421 KB)
[v2] Tue, 16 Sep 2025 06:48:12 UTC (7,940 KB)
[v3] Tue, 17 Nov 2026 13:20:58 UTC (7,936 KB)

来源:arXiv cs.CL · arxiv.org