In the ever-evolving cybersecurity landscape, detecting unseen, zero-day attacks is both urgent and paramount. These sophisticated attacks often lack precedent, posing a challenge to conventional machine learning techniques that rely on prior knowledge and training data. This paper endeavors to detect zero-day and unseen cyber attacks using zero-shot machine learning technique, which holds the promise of identifying these attacks without any prior exposure. This work explores the effectiveness of unsupervised learning in zero-day attack detection. The experimental results demonstrate that autoencoders can identify anomalies in data, which are typically associated with zero-day attacks. When compared with other unsupervised and semi-supervised learning methods, the proposed autoencoder algorithm outperforms its competitors and achieves an accuracy of 99.9%, shedding light on its relative effectiveness in zero-day attack detection.