The Synthetic Web: Single Adversarial Search Result Collapses GPT-5 Accuracy from 65.1% to 18.2%

arXiv·high signal

arXiv 2603.00801 introduces the Synthetic Web Benchmark — procedurally generated adversarial mini-internets that simulate misleading content ranked prominently in search results. Despite unlimited access to truthful sources, a single false source in the top position triggers catastrophic failures across GPT-5, o3, o1, and 4o; agents show minimal search escalation and severe miscalibration. Critical finding for anyone deploying web-browsing or RAG agents in production: adversarial ranking is an unsolved reliability crisis, not an edge case.