The proliferation of IoT systems has introduced significant challenges related to security, scalability, and real-time decision-making. While AI-based techniques, particularly those relying on Deep Learning, have shown strong capabilities for anomaly detection, they remain vulnerable to adversarial attacks and require substantial computational resources for retraining. To address these limitations, we propose a secure and scalable anomaly detection framework that combines AI with Blockchain technology. The proposed approach integrates two complementary modules governed by a unified global detection policy: one for proactive device monitoring and another for real-time anomaly detection. A hybrid Off-chain/On-chain strategy is employed to optimize performance-training and modelling are performed off-chain, while detection decisions are executed on-chain through smart contracts. This design ensures transparent, decentralized, and verifiable outcomes. Experimental validation conducted using the IBRL dataset demonstrates that our framework achieves high detection accuracy, reduces latency and energy consumption, and outperforms existing methods in terms of reliability and computational efficiency.