Needless to say, applications in cybersecurity are increasingly utilizing machine learning (ML) and deep learning (DL) models, thus demanding solutions that offer a trade-off between accuracy, robustness, and computational efficiency. Herein, we study and compare three leading sequential models, i.e., Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM) networks, and Liquid Neural Networks (LiquidNet) in terms of static training performance, dynamic testing robustness, and computational efficiency against the benchmark datasets NSL-KDD and KDD-CUP99. In static training on the NSL-KDD dataset, the performance of LiquidNet was noted at an accuracy of 98.77%, higher than the RNN (98.71%) and close to that of LSTM (99.21%). In the case of training on KDD-CUP99, LiquidNet outperformed RNN and LSTM, attaining 99.60% as opposed to 99.38% and 99.36%, respectively. Further, in dynamic testing, a greater emphasis was noted on the strengths of LiquidNet, having achieved dynamic accuracies of 95.7% using NSL-KDD and 98.6% KDD-CUP99 against RNN (84% and 92%) and LSTM (78% and 90.8%). LiquidNet also holds the edge over RNN and LSTM regarding computational efficiency, with inference times almost on the border of 0 s. In terms of noise robustness, LiquidNet outperformed RNN and LSTM as it showed better tolerance with high noise, thereby assuring higher accuracy and F1-scores across the noise conditions. This research presents a way forward in studying sequential networks’ accuracy, robustness, and efficiency trade-off, identifying LiquidNet as a promising candidate for deployment in time-sensitive scenarios and environments that adverse effects from noise.