Document classification is an essential task for many Cyber Threat Intelligence (CTI) applications that source opensource documents to identify cyber threat patterns such as Indicators of Compromise (IoC), TTP (tactics, techniques, and procedures), Named Entities and more. Traditional machine learning systems require a large number of labelled datasets to automate the document classification process, which is not suitable in the ever-changing CTI landscape. The recent advancements in various Pre-trained Language Models (PLMs) have motivated the popularity of zero-shot text classification approaches. However, a less-studied problem is how such zero-shot methods can be effectively applied in the cybersecurity domain and how their performance is on the document classification task. In this paper, we introduce a self-training approach named Latent Space Refinement (LSR) for zero-shot document classification for CTI applications. Experiments show that our LSR-based method outperforms the relevant state-of-the-art baselines on various scopes of document classification tasks significantly.