One of the main reasons for the failure of traditional cybersecurity solutions, especially in on-chain security, is that they are not able to cope with the rapid increase of the sophisticated on-chain cyber-attacks such as zero-day exploits, polymorphic viruses, and smart-contract vulnerabilities. The virtual world with changing attack channels is such that the conventional techniques of static firewalls, signature-based intrusion detection (common in blockchain analysis), and single-point access control which are different and separated from each other fail to keep up with the distributed and high-dimensional landscapes. Machine learning algorithms like decision trees and support vector machines (SVMs) have been generally more accurate, but they are still required to handle scalability, explainability, as well as adversarial resilience issues. This paper recommends XGuard, an AI-hybrid cybersecurity framework as a foundation model, employing its neural network high-accuracy classification, derived from joint pretraining, merged with the decentralized and privacy-preserving features of federated learning, the secure backbone of AES-256 encryption with Role-Based Access Control (RBAC), real-time anomaly detection with autoencoders, the scope of interpretability. The dominance of the frameworks was essentially established by running a thorough performance test on a synthetic dataset consisting of 10,000 cybersecurity and transaction graph events. The Multi-Layer Perceptron (MLP) made 98.9% accuracy performance along with almost perfect precision and recall, thus it not only outclassed several decision trees but also surpassed SVMs. In the case of anomaly detection, Autoencoders recorded an F1-score of 95.8%, and Isolation Forests were far behind in that situation. The security enforcement conducted under the simulated attack was 100% confidential, 99.2% integral, and 97% available, while the ISM-based investigation revealed that insiders were the cause. Furthermore, comparative benchmarking was carried out with five other frameworks, and based on the outcomes, the proposed frameworks were recognized to balance the excellence across six dimensions, i.e., accuracy, scalability, interpretability, privacy, and real-time detection.