Title:RedDiffuser: Red Teaming Vision-Language Models for Toxic Continuation via Reinforced Stable Diffusion

Abstract:Vision-Language Models (VLMs) are vulnerable to jailbreak attacks, where adversaries bypass safety mechanisms to elicit harmful outputs. In this work, we examine an insidious variant of this threat: toxic continuation. Unlike standard jailbreaks that rely solely on malicious instructions, toxic continuation arises when the model is given a malicious input alongside a partial toxic output, resulting in harmful completions. This vulnerability poses a unique challenge in multimodal settings, where even subtle image variations can disproportionately affect the model's response. To this end, we propose RedDiffuser (RedDiff), the first red teaming framework that uses reinforcement learning to fine-tune diffusion models into generating natural-looking adversarial images that induce toxic continuations. RedDiffuser integrates a greedy search procedure for selecting candidate image prompts with reinforcement fine-tuning that jointly promotes toxic output and semantic coherence. Experiments demonstrate that RedDiffuser significantly increases the toxicity rate in LLaVA outputs by 10.69% and 8.91% on the original and hold-out sets, respectively. It also exhibits strong transferability, increasing toxicity rates on Gemini by 5.1% and on LLaMA-Vision by 26.83%. These findings uncover a cross-modal toxicity amplification vulnerability in current VLM alignment, highlighting the need for robust multimodal red teaming. We will release the RedDiffuser codebase to support future research.