Adversarial Confusion Attack: Disrupting Multimodal Large Language Models

We introduce the adversarial confusion attack as a new class of threats against multimodal large language models (MLLMs).Unlike jailbreaks or targeted misclassification, the goal is to induce systematic disruption that makes the model generate incoherent or confidently incorrect outputs.Practical applications include embedding such adversarial images into websites to prevent MLLM-powered AI Agents from operating reliably.The proposed attack maximizes next-token entropy using a small ensemble of open-source MLLMs.In the white-box setting, we show that a single adversarial image can disrupt all models in the ensemble, both in the full-image and CAPTCHA-style adversarial patch settings.Despite relying on a basic adversarial technique, such as projected gradient descent (PGD), the attack generates perturbations that transfer to both unseen open-source (e.g., Qwen3-VL) and proprietary (e.g., GPT-5.1)models.• We introduce the adversarial confusion attack, which maximizes output entropy to destabilize decoding, and we characterize five distinct modes of resulting model failure.• In the white-box setting, we show that a single perturbation disrupts all models in the ensemble, in both the full-image and Adversarial CAPTCHA settings.