一作ACL 2026Poster· Natural Language Processing
Activation-Guided Local Editing for Jailbreaking Attacks
Jiecong Wang · Haoran Li · Hao Peng · Ziqian Zeng · Zihao Wang · Haohua Du · Zhengtao Yu
摘要
As Large Language Models (LLMs) become indispensable assistants, they remain vulnerable to misuse. Jailbreaking is an essential adversarial technique for red-teaming models to uncover and patch security flaws. However, existing jailbreak methods suffer from significant limitations. Token-level ja...