The authors first construct a Safety Buffer from the poisoned released model: for harmful inputs, they pair them with the released model's safe response and a benign version of the input.
LookAhead Defense uses the model's own safe behavior to block dormant poisoning.
It penalizes only updates that weaken safety, reducing harmful agent actions from 93.9% to 14.2%.
Academic
Wonjun Lee · Vaidehi Patil · Zaid Khan · Yue Zhang · Hyunji Lee · Suhyun Kim · +1 more
UNC Chapel Hill · Kyung Hee University
Research Digest··3 min read
Thread:Agent Security & Attacks
The authors show that dormant poisoning attacks can hide multiple harmful behaviors in language and multimodal agents that activate after benign fine-tuning.
§