Secure and trustworthy energy systems: A four-layer threat model and defense-in-depth framework

The integration of artificial intelligence (AI) is fundamentally reshaping energy systems, offering unprecedented capabilities for renewable integration and autonomous operation. However, this transition introduces complex attack surfaces spanning data, model, system, and physical layers. This paper addresses this critical security paradox by reviewing the inherent vulnerabilities of frontier AI paradigms, including deep learning, large language models (LLM), and reinforcement learning (RL) within the energy context. Building on this, we propose a four-layer threat model that goes beyond established risks to analyze how modern paradigms create novel attack vectors. Specifically, we examine how LLMs enable prompt injection and tool abuse that can manipulate physical processes, how RL agents face reward hacking, and how complex data ecosystems are susceptible to sophisticated poisoning attacks targeting foundational knowledge bases. This paper synthesizes cross-disciplinary evidence and provides a structured threat-modeling and defense-in-depth blueprint applicable to both conventional energy AI pipelines (e.g., forecasting and optimization) and emerging LLM-based operational assistants. To counter these evolving threats, we establish a closed-loop defense-in-depth framework that organizes mitigation measures into four pillars: technical protection, data privacy, system architecture, and governance. We map specific defense technologies to corresponding threat layers to provide a practical protection strategy. Looking forward, we envision pathways toward next-generation trustworthy energy AI, highlighting frontier research directions such as online adaptive defense, verifiable AI, and hardware-enhanced confidential computing. We specifically propose the development of red–blue adversarial training and deception defense systems. Ultimately, this paper advocates for deep interdisciplinary collaboration to ensure AI serves as a reliable driver for decarbonization rather than a security liability.