DUMA-Bench lifts agent attack success to 41.1%

DUMA-Bench evaluated 14 models from five model families across eight domains and found that adding dual-control interaction raised attack success from 26.9% to 41.1%. The authors frame eight vulnerability classes as a test of the model, user, and shared environment together, not model behavior in isolation. Agent teams should add adversarial user-and-environment runs to their eval plans alongside the governance case for agent evaluation; a benchmark pass without interaction testing can overstate deployment safety.