Reproduce “How Much Can We Trust LLM Search Agents? Measuring Endorsement Vulnerability to Web Content Manipulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yimeng Chen, Zhe Ren, Firas Laakom, Yu Li, Dandan Guo, Jürgen Schmidhuber
The authors introduce SearchGEO to test whether search agents endorse manipulated results. It inserts attacks among cached real search results in 44 queries and evaluates 13 backends. Attack-only endorsement rates range from 0.0% to 31.4%. On Gemini-3-Flash, Mode 2B endorsement rises from 39% to 77% as distinct sources increase from one to three; these findings do not estimate real-world prevalence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.