The recent era has witnessed a huge development in Internet of Things (IoT) devices integrated with cloud services, amplifying significant security vulnerabilities. Such threats include unauthorized access, identity spoofing, data leakage, and network-based attacks like spoofing, Distributed Denial of Service (DDoS), and man-in-the-middle (MITM) attack. According to latest forecast, a significant amount of IoT connected devices around 41 billion are expected which result in escalating the complexity and volume of data transfer between resource constrained devices especially through cloud platforms. The problem even more multifaceted from recent cybersecurity report which indicates that approximately 57% of IoT devices are currently venerable to attack for medium to high severity. These vulnerabilities contributed to a very high impact in IoT target cyber incidents, and even some reports indicate increase of 40% exceeding year by year. This paper proposes an AI-enhanced adaptive security framework that attempts to integrate lightweight edge processing with cloud-hosted machine learning models to provide scalable, real-time protection. The framework incorporates preventive mechanisms (e.g., device authentication), detective analytics using ML-based anomaly detection, and responsive actions like session termination and automated rollback to mitigate evolving threats, the harmful traffic automatically triggers recovery actions such as IP blocking, session termination, or resource rollback. Practical evaluation on Raspberry Pi edge devices and AWS cloud services shows a result of 98.9% detection rate, 1.2% false positives, 1.1% false negatives, <120 ms classification latency, and <200 ms end-to-end mitigation latency. The system supported 50 IoT devices with moderate resource use and improved detection of low-frequency attacks through retraining. These results validate the feasibility of integrating adaptive AI, lightweight edge-cloud architectures, and automated recovery for practical IoT-cloud security.