The digitalization of power systems has introduced cyber-physical vulnerabilities in smart grid infrastructures that traditional, rule-based detection mechanisms are insufficient to address. This paper presents an AI-enhanced cybersecurity risk assessment framework that integrates the NIST Cybersecurity Framework (CSF~2.0), NIST~SP~800-30, and MITRE~ATT\&CK for ICS with machine learning-based threat detection. Three classifiers: Logistic Regression, Random Forest, and XGBoost --- were trained and evaluated on the Blockchain-Enabled 6G Microgrid Dataset, a multi-layer dataset comprising physical, communication, and blockchain operational features with a severe class imbalance ratio of 29.5:1. Model performance was assessed using accuracy, precision, recall, F1-score, and ROC-AUC, with recall and F1-score prioritized given the asymmetric cost of false negatives in cybersecurity contexts. XGBoost achieved the highest performance (accuracy: 0.993, F1-score: 0.885, ROC-AUC: 0.974), outperforming Logistic Regression and Random Forest through its sequential boosting architecture and scale_pos_weight parameter, which up-weighted the minority attack class during training. A key contribution is the systematic mapping of XGBoost probabilistic output scores to NIST SP800-30 likelihood tiers, enabling continuous quantitative risk scoring via R = L x I and supporting automated protective responses in real time. A case study of the 2015--2016 Ukrainian power grid attacks further contextualizes the framework's applicability to real-world critical infrastructure scenarios.