Evaluating Deep-Search Agents under Hierarchical Web Evidence Poisoning

This study demonstrates that deep-search AI agents can be manipulated through poisoned web evidence, exposing a real vulnerability in agentic systems that rely on untrusted online sources.

Auto-scored: The paper documents a concrete attack vector (hierarchical web evidence poisoning) that can mislead deep-search agents, representing a real escalation in agentic AI misuse risk, while also providing a partial defensive evaluation.