The rise of machine unlearning marks a significant shift in computational strategies, focusing on deliberate mechanisms to erase specific data from trained models, ensuring compliance with rigorous privacy regulations. Among these is the ‘right to be forgotten,’ which allows the removal of certain data or classes from machine learning models. Retraining models from scratch is impractical in most cases. Prior studies often rely on subsets of training data or stored metadata to adjust model weights for unlearning. Yet, strict regulations often demand swift data deletion, leaving no access to training samples for unlearning. This raises the question: Can we unlearn a single class or multiple classes from a machine learning model without access to the original training data? In this paper, we address the zero-shot machine unlearning scenario, where no original data samples are available. We propose Z-MUGAN, Zero-shot Machine Unlearning Using GANs, a novel approach that leverages model inversion to generate a proxy of the original training data. This proxy is subsequently used to update the original model through impair and repair steps. Empirical evaluations across various benchmarks confirm that our approach achieves state-of-the-art performance under the rigorous constraints of zero-shot unlearning.