Multimodal Large Language Models (MLLMs) are susceptible to attacks based on Jailbreak Probability, known as Jailbreak-Probability-based Attacks (JPA). These attacks utilize a specialized Jailbreak Probability Prediction Network (JPPN) to detect and refine adversarial perturbations within input images. This optimization process aims to maximize the likelihood of generating harmful responses from the MLLM, even when employing minimal perturbation bounds and limited iterations. The attack functions by altering the hidden states of the input image within the MLLM, thereby increasing the predicted jailbreak probability. Mitigation steps: **For AI Developers:** * [Integrate Jailbreak-Probability-based Defensive Noise (JPDN) as a pre-processing step to input images, aiming to lower their jailbreak probability.] **For Model Trainers/Fine-tuners:** * [Apply Jailbreak-Probability-based Finetuning (JPF) to adjust the model's parameters, thereby decreasing the predicted jailbreak probability of adversarial images.]