DarkLLM: Learning Language-Driven Adversarial Attacks with Large Language Models

Researchers propose the DarkLLM attack framework, which trains a large language model to convert natural language attack instructions into latent attack vectors, then decode them into visual adversarial perturbations. Leveraging natural language instruction fine-tuning, the framework unifies targeted, untargeted, segmentation, and multi-model attacks within a single framework, with each instruction able to generate perturbations that induce desired behavior on heterogeneous models. In experiments covering 4 tasks, 13 datasets, and 15 models, DarkLLM with only 1B parameters can follow attacker instructions and generate highly effective attacks against CLIP, SAM, and frontier large language models; the authors state this reveals systematic weaknesses in modern foundation models.