1 paper
Wenkai Shen, Pengyang Zhou, Jiahe Xu +5
LLM-powered search agents enable multi-step reasoning and tool use. However, these capabilities introduce retrieval-induced safety degradation, as harmful intents may decompose int…