With the advance in optimal control and reinforcement learning, latest controllers have demonstrated exceptional performance on complex multi-limbed robots, including the application in aerial manipulation, quadrupedal locomotion, and dexterous hand skills. Despite these advancements, comprehensive safety validation remains a prerequisite for their large-scale real-world deployment. Current state-of-the-art (SOTA) controllers exhibit robustness against standard testing paradigms, domain randomization, and evaluations by human experts. In this talk, I will first introduce our progress on enhancing dexterity of the robots, and then our recent research revealing vulnerabilities in long-tested state-of-the-art controllers on the legged locomotion and dexterous manipulation when subjected to minor perturbations generated by AI agents. This study highlights critical safety concerns and emphasizes the necessity of addressing these vulnerabilities to enhance system reliability. The failure cases identified in our analysis offer valuable insights into system components, providing a foundation for improving the robustness and safety of black-box neural controllers.