API-protected LLMs can leak proprietary information due to a softmax bottleneck, revealing hidden model details.