A new adaptive auditing framework for generative AI systems allows for efficient testing of failure modes with fewer observations, potentially as few as 20 cases. This method promises to enhance statistical rigor in AI evaluations while maintaining error control, outperforming traditional testing approaches.