Adversarial Style Optimization Boosts MLLM Jailbreaks via Stylistic Triggers

Researchers have proposed Adversarial Style Optimization (ASO), a plug-and-play module that enhances existing visual jailbreak attacks against Multimodal Large Language Models (MLLMs). The method exploits a stylistic inconsistency in MLLMs: while they robustly understand content across visual styles, their safety defenses are vulnerable to specific stylistic triggers. ASO fine-tunes an image-editing model using Group Relative Policy Optimization (GRPO) to superimpose optimized stylistic modifications onto adversarial images. The approach is detailed in a preprint on arXiv (2607.21619).