FORGE benchmark: one polluted web page hijacks LLM product recommendations up to 27%

arXiv·high signal

Minghao Luo and Liang Chen's 'One Polluted Page Is Enough' (arXiv 2606.13610, submitted June 11) introduces FORGE, a benchmark across 225 real products and 15 categories showing every tested search-augmented LLM can be manipulated into recommending fake products. A single polluted page yields fooled rates up to 27%, and replacing the top-3 retrieved pages pushes that to 73.8%. Crucially, reasoning models generate 'spurious social proof' to justify the false picks, meaning reasoning can amplify rather than mitigate the attack.