Abstract <p>The advancement of cyberattacks has sidelined traditional security measures to secure a dynamic environment. Zero Trust (ZT) provides a reliable framework to secure organizations. One of the key aspects of ZT is the threat detection. It is used for timely mitigation and maintain proactive cyber-posture. However, many studies lack implementation and architectural details for integrating it with ZT. Focusing only on accuracy often creates complex models. Various existing models operate on discrete time steps, which limits the ability to accurately detect real-time attack patterns. In addition, various models require fine-tuning when input distributions change. It limits their adaptability. To address it, a Liquid Neural Network (LNN)-based ZT architecture called Zero Adaptive Trust (ZAT) has been proposed. It uses continuous-time dynamics to improve threat detection. Two liquid layers with different time constants are used to improve temporal awareness during detection. A two-phase training is performed with a dual learning rate control system for stable and adaptive training. To evaluate the robustness of ZAT, three variations of the CICIDS2017 dataset (based on the class distribution) are used to reflect real-life scenarios. The best performance was achieved with the slightly imbalanced variant, supported with various analyses. It achieved convergence in the least epochs and time, with the various other near-ideal metrics. ZAT is specifically designed to handle access requests with low inference time and minimal trade-off in decision accuracy. This study provides a foundation for the advancement of ZT and adaptive threat detection to strengthen existing approaches. Its flexible design enables extension to other domains with minimal alterations.</p> Graphical Abstract <p></p>