The rapid proliferation of Internet of Things (IoT) devices across industrial, domestic, and critical-infrastructure environments has created a correspondingly rapid expansion of cyber-attack surfaces, while the heterogeneity, resource constraints, and high-volume dynamic traffic of IoT networks render conventional signature- and rule-based security mechanisms increasingly inadequate. This article presents a comprehensive review of twelve recent studies (2020–2026) on artificial intelligence (AI)-driven approaches to securing IoT networks, spanning machine learning and deep learning intrusion detection systems (IDS), federated and blockchain-assisted learning frameworks, 5G software-defined network (SDN) security architectures, service-oriented edge-layer architectures, and large-scale systematic and taxonomic reviews. Reported detection accuracies across the empirical studies range from approximately 82% to 98.25%, with federated and hybrid blockchain–AI approaches offering the strongest privacy–accuracy trade-offs and 5G-SDN frameworks achieving the lowest detection latencies (as low as 15 milliseconds). Despite these advances, the literature consistently reports unresolved research gaps, including limited adaptability to evolving traffic patterns, elevated false-positive rates, insufficient attention to the computational and energy constraints of IoT endpoints, dependence on single machine learning or deep learning algorithms rather than integrated pipelines, and an absence of unified, lightweight, and quantum-resistant frameworks validated on real hardware. Building on this synthesis, the review articulates a research problem, four research objectives, and a proposed seven-stage AI-driven security framework — spanning data collection, preprocessing, feature extraction and selection, AI-based threat detection, attack classification, adaptive security response, and continuous monitoring — intended for validation on benchmark datasets (CICIoT2023, TON_IoT, BoT-IoT/N-BaIoT) and on constrained hardware platforms (Raspberry Pi, ESP32). The proposed research direction aims to deliver a lightweight, adaptive, and empirically validated AI-driven security solution suited to resource-constrained IoT deployments.