A study posted on arXiv reports that a single strategically placed webpage can influence answers from web-searching AI systems. The researchers tested ChatGPT, Gemini and Sonar Pro with prompts on topics where reliable online information was scarce, varying prompt tone and other attacker-controlled factors.
Across the tested conditions, content controlled by an attacker appeared in responses 31% of the time. Skeptical prompts slightly reduced success and increased warnings, but the study found no large effects from prompt tone or model choice. The paper is an extended version slated to appear at the NDSS Symposium 2027.
Why it matters
People using these systems for information on topics with few reliable sources may encounter strategically published material, while the study found that skeptical wording offered only limited protection.
