Vulnerabilities and Risk Analysis of Multi-Agentic AI-RAG in Autonomous Vehicle Reinforcement Learning Frameworks: Safety Mitigation in Training, Simulation, and Real-World Testing

Modern CI/CD DevOp practices are continually evolving with the integration of AI-enabled components, rapid advancements in Large Language Models (LLMs), Retrieval-Augmented Generation (RAG) systems, and agentic AI frameworks. These AI-enabled components enhance automation in real time, perform speed testing, and enable autonomous decision making across software lifecycles. However, their increasing autonomy and data-driven nature also expose different classes of vulnerabilities that traditional security measures fail to address. Agentic and multi-agent autonomous systems ingrained with Reinforcement Learning Methods may exhibit signs of data, model, and memory poisoning and degrade with time. This paper presents a comprehensive overview of the emerging security risks in AI-augmented and DevOps environments, including prompt injection, data poisoning, agent manipulation, and reinforcement learning failures. This study aims to provide guidance to practitioners and researchers for developing secure, resilient, and trustworthy AI integrated AV and self-driving capability systems.