Cybersecure by Design: Managing Adversarial Risk and Resilience in AI-Enabled Critical Infrastructure

Abstract

The use of artificial intelligence in critical infrastructure creates a new class of cyber-physical and institutional risks. AI systems depend on extensive data pipelines, complex models, software supply chains, cloud services, interfaces, and continuous updates. These features expand the attack surface beyond traditional information technology and expose models to data poisoning, evasion, model extraction, inference attacks, manipulation of sensors, and unsafe automated responses. This review examines cybersecurity as a central component of AI governance rather than a separate technical speciality. It draws on the thematic synthesis of 95 high-quality studies and frameworks covering public administration, digital services, smart cities, and infrastructure-related applications. The evidence indicates that many governance frameworks acknowledge privacy and ethics but provide limited operational guidance for threat modelling, adversarial testing, incident response, resilience, and audit evidence. The review presents a lifecycle approach to AI security that spans design, data preparation, model development, validation, deployment, monitoring, change management, and retirement. It also explains how risk tiering can connect system impact to proportionate controls and how assurance records can support accountability. Particular attention is given to smart grids, transport, healthcare, buildings, and structural monitoring, where security failures can move quickly from digital disruption to physical or social harm. The article concludes that cybersecure AI requires integrated governance, continuous testing, and recovery planning from the beginning of system design.