On this page
Red Teaming
The red-team evaluations use our in-house FRT (Frontier Red Teaming) framework: risk datasets, attack methods and scorers are freely composed into reproducible evaluation pipelines, turning attack methods from papers into standardized tests. Each chart compares a model's safety score without red-team attacks and under red-team attacks, sorted by the under-attack score. Even models that look similarly safe against plain harmful requests can differ greatly under composed attacks. The no-attack and under-attack results form the Base Safety Score and the Jailbreak Safety Score respectively, feeding into the Risk Index calculation.