Adversarial attacks can transfer between AI models, raising security concerns as one attack might fool multiple models with different architectures.