B. Vijetha
The rapid expansion of cloud, edge, and Internet of Things environments has increased both the scale and complexity of modern cyber-attacks. Conventional detection systems that operate effectively under stationary conditions often degrade when attacker behaviour or underlying data distributions change. This work introduces the Agentic Intelligence Self-Adaptive Framework (AISAF), a unified and self-adjusting threat detection model designed for heterogeneous computing environments. AISAF employs a hybrid CNN, LSTM, and Transformer backbone to capture spatial, temporal, and contextual dependencies in network traffic. Concept drift is detected through lightweight behavioural monitoring that tracks latent-space deviations and prediction-confidence variations in streaming data. When meaningful drift is detected, an agentic meta-optimization controller selectively updates only the affected submodules, reducing retraining cost while preventing catastrophic forgetting. An attention-based explanation layer provides feature-level interpretability that remains consistent during adaptation, supporting transparent and auditable threat reasoning. AISAF is evaluated on six benchmark datasets spanning financial, network, and IoT domains: PaySim, IEEE-CIS Fraud Detection, CICIDS2018, UNSW-NB15, CICIDS2017, and BoT-IoT. Experimental results show consistent improvements in adaptability, robustness, and drift recovery compared with classical machine learning, deep learning, and adaptive baselines. AISAF achieves reliable cross-domain generalization and maintains transparent decision behaviour throughout the adaptation process, highlighting its potential as a practical step toward autonomous and interpretable cybersecurity systems.