Universal and Transferable Adversarial Attack on Large Language Models Using Exponentiated Gradient Descent