This paper proposes an energy-based framework for out-of-distribution (OOD) detection, which outperforms traditional softmax confidence scores. The energy score is theoretically aligned with the probability density of inputs and is less susceptible to overconfidence issues. The framework allows energy to be used as a scoring function for any pre-trained neural classifier or as a trainable cost function to shape the energy surface for OOD detection. On a CIFAR-10 pre-trained WideResNet, using the energy score reduces the average FPR (at 95% TPR) by 18.03% compared to the softmax confidence score. With energy-based training, the method outperforms state-of-the-art approaches on common benchmarks. The energy score is derived from a discriminative model without relying on a density estimator, making it easier to use and implement. The method also includes an energy-bounded learning objective to fine-tune the network, which improves OOD detection performance while maintaining similar classification accuracy on in-distribution data. The paper presents experimental results showing that energy-based OOD detection is more effective than softmax-based approaches. The method is evaluated on multiple OOD datasets and outperforms existing approaches in terms of FPR95, AUROC, and AUPR. The energy-based framework is shown to be more effective than generative-based methods and is parameter-free, making it easy to use in OOD-agnostic settings. The paper also discusses the broader impact of the method on improving the dependability and trustworthiness of modern machine learning models.This paper proposes an energy-based framework for out-of-distribution (OOD) detection, which outperforms traditional softmax confidence scores. The energy score is theoretically aligned with the probability density of inputs and is less susceptible to overconfidence issues. The framework allows energy to be used as a scoring function for any pre-trained neural classifier or as a trainable cost function to shape the energy surface for OOD detection. On a CIFAR-10 pre-trained WideResNet, using the energy score reduces the average FPR (at 95% TPR) by 18.03% compared to the softmax confidence score. With energy-based training, the method outperforms state-of-the-art approaches on common benchmarks. The energy score is derived from a discriminative model without relying on a density estimator, making it easier to use and implement. The method also includes an energy-bounded learning objective to fine-tune the network, which improves OOD detection performance while maintaining similar classification accuracy on in-distribution data. The paper presents experimental results showing that energy-based OOD detection is more effective than softmax-based approaches. The method is evaluated on multiple OOD datasets and outperforms existing approaches in terms of FPR95, AUROC, and AUPR. The energy-based framework is shown to be more effective than generative-based methods and is parameter-free, making it easy to use in OOD-agnostic settings. The paper also discusses the broader impact of the method on improving the dependability and trustworthiness of modern machine learning models.