Cyberattacks on critical infrastructure, such as energy, water, and transportation systems, are becoming more frequent and complex, and existing defenses struggle against new or unknown threats. Key AI challenges include system heterogeneity, overreliance on data over physical models, sensitivity to added security demands, and uncertainty in linked AI models. A proposed solution, CPSAgentic, integrates physics-informed learning and dynamic knowledge graphs with AI agents to enable resilient, trusted, and adaptive responses to anomalies and faults.

While Large Language Models (LLMs) show potential for tackling security challenges, they often fail to detect rare or abnormal attack behaviors. As these issues become more complex, LLMs will need to be designed specifically for security needs. However, fragmented LLM ecosystems and diverse requirements prevent proper integration of these models. This project proposed creating a foundational, security-focused LLM and high-quality datasets to enable effective fine-tuning for protecting critical infrastructure.