作者位次 #5/5ACL 2026Findings· Findings of ACL
T raj G uard: Streaming Hidden-state Trajectory Detection for Decoding-time Jailbreak Defense
Cheng Liu · Xiaolei Liu · Xingyu Li · Bangzhou Xin · Kangyi Ding
摘要
Existing jailbreak defense paradigms primarily rely on static detection of prompts, outputs, or internal states, often neglecting the dynamic evolution of risk during decoding. This oversight leaves risk signals embedded in decoding trajectories underutilized, constituting a critical blind spot i...