1 paper
Cheng Liu, Xiaolei Liu, Xingyu Li +2
Existing jailbreak defense paradigms primarily rely on static detection of prompts, outputs, or internal states, often neglecting the dynamic evolution of risk during decoding. Thi…