Tag
The paper introduces TA-SPA, a black-box jailbreak attack framework for multimodal large language models that uses text-anchored semantic perturbations to achieve effective and transferable attacks against safety alignments.