Attackers Can Poison AI Research Agents Using Reddit and Wikipedia Content

Cornell Tech research shows attackers can poison web-retrieval AI research agents by appending short, believable promotional snippets to high-ranking user-generated content (Reddit, Wikipedia, forums). Experiments demonstrate substantial influence—snippet attacks produced 38–51% conditional mention rates and full-page injections led to 30–53% repetition—while standard defenses (blocking UGC, filtering, similarity checks) prove ineffective or damaging to answer quality.