Cyber threats have evolved significantly, rendering conventional anomaly-based and signature-based detection techniques ineffective against sophisticated as well as adaptive cyberattacks. To address these limitations, we propose an AI-driven cybersecurity model using Reinforcement Learning (RL) to detect, prevent, as well as mitigate cyber threats in real-time. Unlike conventional ML approaches that rely on static models, RL dynamically learns attack patterns, adapts to new threats, and enhances defense mechanisms autonomously. The proposed model utilizes Deep Q-Networks (DQN) and policy gradient methods to analyze network traffic, detect anomalies, as well as respond to cyber threats with minimal false positives. We evaluate the model on benchmark datasets like NSL-KDD and CIC-IDS2017, comparing its performance against traditional Intrusion Detection Systems (IDS) and supervised ML models like random forest (RF), Support Vector Machines (SVM), as well as Deep Neural Networks (DNNs). Our experimental findings demonstrate that RL-based system achieves higher detection accuracy (above 95%), reduces false alarms, and effectively mitigates cyber threats. In addition, we present a federated RL method to leverage threat intelligence among different federated networks in a privacy-preserving way. According to these findings, RL has the possibility to revolutionize cybersecurity by being an adaptive, intelligent and real-time defence mechanism.