This study presents an analytical model for reporting high-impact crimes in Colombian municipalities using data mining and machine learning techniques. The CRISP-DM methodology guided the integration of official data from law enforcement, judicial, and demographic sources. K-means clustering was used to identify geographic crime patterns, and a multilayer perceptron neural network was applied to classify municipalities into High, Medium, or Low crime levels. The neural network achieved an accuracy of 91.8%, outperforming Random Forest, Naive Bayes, and Logistic Regression models. The model’s performance, reflected in a weighted F1-score of 0.92, demonstrates high precision in identifying municipalities with both high and low crime incidence. Principal Component Analysis (PCA) revealed two latent dimensions: PC1 (general criminal activity) and PC2 (organized and severe crime), offering a multidimensional understanding of crime dynamics. By incorporating normalized crime rates and procedural variables, the model enhances the predictive accuracy and supports the design of territorially grounded, evidence-based public security policies.