Diffusion Language Models for Perplexity-Aware Black-Box Attacks