The rapid growth of digital infrastructures and interconnected systems has significantly increased the surface for cyber attacks, making accurate and automated attack classification a critical research challenge. This paper aims to build a foundation to address this industry issue through the proposal, implementation, and testing of a comparative study of different machine learning algorithms to determine the optimal model for classifying and categorizing cyber attacks. This study utilizes a curated, pre-processed attack surface dataset containing a representative, and multi-dimensional, sample of class-attack scenarios, associated tools, and mitigation strategies. Pre-processing activities on the dataset included the incorporation of robust imputation techniques for missing values, inter- and intra-class data distributions for the presence of imbalances, the presence of appropriate scaling of features, and the incorporation of an appropriate noise signal. Under a common framework, six classifiers were compared. These included two widely used ensembles, simple Gaussian Naive Bayes, and a Soft Voting Ensemble. The findings demonstrate that tree-based classifiers achieve strong and consistent performance for multi-class cyber attack classification. Furthermore, the results indicate that probabilistic and ensemble voting classifiers exhibit comparatively lower effectiveness under the same experimental conditions. In a practical sense, these findings demonstrate the potential of using a model-agnostic, tree-based classifier to tackle complex, high-dimensional multi-class classification problems, particularly within the context of cyber attack surveillance and activity mitigation.