In the field of cybersecurity, the continuous evolution of user behavior and attackers’ tactics leads to dynamic changes in data distribution over time, resulting in a decline in the performance of security detection models. This paper introduces an innovative continual drift detection and adaption cybersecurity framework called CDDA. CDDA combines contrastive learning, unsupervised clustering, and uncertainty sampling to effectively optimize the embedding space, accurately detect malicious samples, and efficiently sample while reducing the labeling burden on security analysts. The experiments are conducted on two publicly available multi-temporal datasets - APIGraph and Kyoto. The experimental results show that CDDA significantly reduces the false negative rate from 26% to 11% and improves the F1 score from 80% to 90%.