Machine learning (ML) models are often vulnerable to performance degradation when encountering Out-of-Distribution (OOD) data that significantly deviates from their training data distribution. This results in degraded performance and underscores the importance of robust OOD detection for real-world applications. Existing OOD detection methods, including recent gradient-based techniques, generally fail to fully utilize the directional information of gradients, which is critical for distinguishing between in-distribution(ID) and OOD data effectively. This paper introduces a novel gradient-based framework that not only addresses these challenges by incorporating the direction of gradients but also significantly advances OOD detection methodologies by introducing a robust scoring mechanism that leverages the entire model’s gradient vectors. Through extensive experimentation and analysis of disruptive patterns in model behaviour, our framework refines detection mechanisms and enhances the robustness of ML models, thereby improving their reliability and generalization across diverse contexts. The experiment results with three benchmark data and comparison with three state-of-the-art OOD detection approaches mark a significant step forward in the field of OOD detection, offering a comprehensive solution to one of the most pressing challenges in machine learning.