Deep learning has become an increasingly common technique for various control\nproblems, such as robotic arm manipulation, robot navigation, and autonomous\nvehicles. However, the downside of using deep neural networks to learn control\npolicies is their opaque nature and the difficulties of validating their\nsafety. As the networks used to obtain state-of-the-art results become\nincreasingly deep and complex, the rules they have learned and how they operate\nbecome more challenging to understand. This presents an issue, since in\nsafety-critical applications the safety of the control policy must be ensured\nto a high confidence level. In this paper, we propose an automated black box\ntesting framework based on adversarial reinforcement learning. The technique\nuses an adversarial agent, whose goal is to degrade the performance of the\ntarget model under test. We test the approach on an autonomous vehicle problem,\nby training an adversarial reinforcement learning agent, which aims to cause a\ndeep neural network-driven autonomous vehicle to collide. Two neural networks\ntrained for autonomous driving are compared, and the results from the testing\nare used to compare the robustness of their learned control policies. We show\nthat the proposed framework is able to find weaknesses in both control policies\nthat were not evident during online testing and therefore, demonstrate a\nsignificant benefit over manual testing methods.\n