Autonomous vehicles (AVs) are already driving on public roads around the US; however, their rate of deployment far outpaces quality assurance and regulatory efforts. Consequently, even the most elementary tasks, such as automated lane keeping, have not been certified for safety, and operations are constrained to narrow domains. First, due to the limitations of worst-case analysis techniques, we hypothesize that new methods must be developed to quantify and bound the risk of AVs. Counterintuitively, the better the performance of the AV under consideration, the harder it is to accurately estimate its risk as failures become rare and difficult to sample. This thesis presents a new estimation procedure and framework that can efficiently evaluate and AV's risk even in the rare event regime. We demonstrate the approach's performance on a variety of AV software stacks. Second, given a framework for AV evaluation, we turn to a related question: how can AV software be efficiently adapted for new or expanded operating conditions? We hypothesize that stochastic search techniques can improve the naive trial-and-error approach commonly used today. One of the most challenging aspects of this task is that proficient driving requires making tradeoffs between performance and safety. Moreover, for novel scenarios or operational domains there may be little data that can be used to understand the behavior of other drivers. To study these challenges we create a low-cost scale platform, simulator, benchmarks, and baseline solutions. Using this testbed, we develop a new population-based self-play method for creating dynamic actors and detail both offline and online procedures for adapting AV components to these conditions. Taken as a whole, this work represents a rigorous approach to the evaluation and improvement of AV software.