大模型对齐与安全

Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models

越狱攻击下的鲁棒有害特征：来自大型语言模型中注意力头特化的机制证据

Yanchen Yin, Dongqi Han, Linghui Li

专题命中 越狱攻击 ：jailbreak(title,abstract)；alignment(abstract)；safety(abstract)；分类 cs.AI

AI总结 通过分析注意力头特化，发现越狱攻击通过抑制早期层的对抗妥协头（ACHs）绕过安全对齐，而中间层的安全对齐头（SAHs）保持鲁棒激活，揭示了鲁棒有害特征现象，并利用这些持久激活实现无需训练的检测。

Comments 33 pages, 19 figures. Accepted at ICML 2026 as an Oral presentation

When the Prompt Becomes Visual: Vision-Centric Jailbreak Attacks for Large Image Editing Models

当提示变为视觉：面向大型图像编辑模型的以视觉为中心的越狱攻击

Jiacheng Hou, Yining Sun, Ruochong Jin, Haochen Han, Fangming Liu, Wai Kin Victor Chan, Alex Jinpeng Wang

专题命中 越狱攻击 ：jailbreak(title,abstract)；safety(abstract)；trustworthy(abstract)；分类 cs.AI

AI总结 提出首个视觉到视觉的越狱攻击VJA，通过纯视觉输入传递恶意指令，并构建安全基准IESBench，在商业模型上攻击成功率高达80.9%，同时提出无需训练的内省多模态推理防御方法。

Comments Accepted for spotlight and oral presentation at ICML 2026 (Project: https://csu-jpg.github.io/vja.github.io/)

MetaBreak: Jailbreaking Online LLM Services via Special Token Manipulation

MetaBreak: 通过特殊标记操纵越狱在线LLM服务

Wentian Zhu, Zhen Xiang, Wei Niu, Le Guan

专题命中 越狱攻击 ：alignment(abstract)；safety(abstract)；jailbreak(abstract)；分类 cs.AI

AI总结 提出利用特殊标记构造攻击原语，绕过LLM安全对齐和内容审核，并发现防御困难，实验显示MetaBreak在内容审核下优于现有方法。

Comments Accepted version. Revised to match the version accepted to the 2026 IEEE Symposium on Security and Privacy (SP); added publication information and DOI

Journal ref Proceedings of the 2026 IEEE Symposium on Security and Privacy (SP), pp. 98-117, IEEE Computer Society, 2026