Data poisoning attacks pose a significant security risk to network security software that utilizes machine learning (ML) for network intrusion detection. As network traffic continues to surge, ML becomes indispensable in detecting and characterizing malicious actors attempting to infiltrate computer networks. However, conventional ML assumes a benign environment, leaving room for adversaries to violate this assumption during the training phase. Detecting data poisoning attacks proves to be a challenging task, as attackers employ subtle alterations in the training data to create backdoors, trojans or triggers. Traditional techniques for addressing data poisoning attacks often focus only on enhancing ML model robustness rather than detecting poisoned data, necessitating the development of novel, more effective approaches. Hence, there is an urgent need to develop new methods for identifying poisoned data, ensuring the security of ML. We introduce a novel approach that harnesses the power of topological data analysis and unsupervised learning, enabling the early identification of poisoned data before training an ML model for network intrusion detection. Leveraging our approach, the extraction of topological features and subsequent application of clustering techniques leads to the creation of new clusters exclusively composed of poisoned data for removal prior to ML model training.