Preprint tests coordinated search-result attacks on AI agents

A new arXiv preprint reports that a controlled search intermediary can steer LLM-based search agents by inserting one result per query and coordinating evidence across an agent's research path. The authors describe their Authority-Chain Hijack method as reaching 55.9% attack-success rate and 83.3% MaxN attack-success rate on the full SafeSearch test split. A trace-guided strategy system reached 71.4% and 95.0%, respectively, in held-out evaluation. The work frames search and page-observation channels as a security boundary rather than only a source of isolated prompt injections.