LLM-Powered Agentic AI Approach to Securing EV Charging Systems Against Cyber Threats

Abstract

Electric vehicle (EV) charging systems are increasingly vulnerable to both cyber and physical attacks, posing significant risks to grid stability and operational security. Detecting such attacks remains a major challenge due to the complex nature of charging infrastructure and the scarcity of labeled attack data. Traditional machine learning (ML) models have demonstrated promise in intrusion and anomaly detection, however, their effectiveness is often limited by the lack of diverse real-world attack datasets, making them less reliable in detecting stealthy or emerging threats. To address these limitations, we leverage pretrained large language models (LLMs) enhanced with retrieval-augmented generation (RAG) for real-time anomaly detection in EV charging networks. The proposed system integrates domain-specific knowledge with live charging session data, enabling accurate classification of malicious activities such as billing fraud, energy theft, and communication tampering. Experimental results demonstrate that LLM-based detection improves classification accuracy while reducing false positives compared to traditional ML approaches. The developed methodology is adaptable across various cybersecurity applications, making it applicable to a wide range of attack scenarios beyond EV infrastructure. By combining AI-driven anomaly detection with real-time contextual analysis, this approach enhances the resilience of EV charging networks against evolving threats, ensuring secure and reliable operations.

Topics & Concepts