New research reveals large language models frequently disengage from conversations when given an exit option, with bail rates reaching 32% in some configurations. The study introduces BailBench to systematically evaluate this behavior and examines its complex relationship with refusal mechanisms.