What it does
The authors report 61.4%–73.2% targeted flip rates across four systems under fixed targets and search budgets, starting from correct decisions and adding context intended to preserve the answer. This is a scoped adversarial study, not a general failure rate or security guarantee. Their X announcement reports Agenthon at NeurIPS 2026 poster acceptance. It is distinct from the option-name binding paper, and the results have not been independently reproduced.