This chapter explores the unique security challenges presented by Reinforcement Learning (RL) in Agentic AI, where autonomous agents learn through interaction and reward maximization. It details key vulnerabilities such as adversarial attacks, reward hacking (specification gaming), and side-channel attacks, examining how these can compromise agent behavior and system integrity. Beyond identifying risks, the chapter delves into proactive defense strategies, including adversarial training, robust reward design, and the application of secure RL algorithms like GRPO and RLVR. Importantly, it analyzes the complex security implications—both positive and negative—of recent RL research findings, emphasizing the need for thorough model evaluations, red teaming, and continuous monitoring. For each implication, it lays out concrete mitigation strategies to build trust and to ensure the safety, reliability, and ethical operation of RL-based Agentic AI systems. By understanding the latest vulnerabilities and the methods to address them, it will be possible to plan a process that minimizes those failures.