Cross-Modal Obfuscation for Jailbreak Attacks on Large Vision-Language Models

Lei Jiang, Zixun Zhang, Zizhou Wang, Xiaobing Sun, Zhen Li, Liangli Zhen, Xiaohua Xu

Keywords: Large Vision-Language Models, Jailbreak attacks, Black-box attack methods

TL;DR: This paper introduces CAMO, a novel black-box jailbreak attack that exploits LVLMs' cross-modal reasoning to stealthily bypass safety mechanisms with high efficiency and strong transferability.

Abstract:

Large Vision-Language Models (LVLMs) demonstrate exceptional performance across multimodal tasks, yet remain vulnerable to jailbreak attacks that bypass safety mechanisms. Existing jailbreak methods suffer from two critical limitations: insufficient stealth against input-level defense filters and high computational costs from lengthy prompts or iterative procedures.
In this work, we present Cross-modal Adversarial Multimodal Obfuscation (CAMO), a black-box jailbreak framework that decomposes harmful instructions into benign-looking textual and visual clues.
CAMO leverages LVLMs' cross-modal reasoning to reconstruct attack intent while each component appears harmless in isolation, evading defense filters.
Our compositional obfuscation design achieves superior efficiency, using only 12.6\% of tokens required by existing methods while achieving high attack success rates of 81.82\% on GPT-4.1-nano and 93.94\% on DeepSeek-R1.
CAMO bypasses multiple defense mechanisms with 100\% evasion rate, demonstrating effectiveness across both open-source and closed-source models. This work exposes critical vulnerabilities in current multimodal safety protocols and underscores the need for more sophisticated defense strategies.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 5630