1
Geometric Configurations of Perturbed Jailbreak Prompts
揭秘大模型安全软肋:扰动型越狱提示的几何结构,小模型内部表征泄露攻击机制。
arXiv:2607.20581v1 Announce Type: cross Abstract: Perturbation techniques that turn unsuccessful jailbreak prompts into successful ones are continuous…