We present a systematic review of 63 high-quality peer-reviewed studies (2020–2024) aimed at AI-based cybersecurity solutions. We propose a new taxonomy that categorizes the identified works into four broad categories: traditional machine learning (SVM, XGBoost), deep learning approaches (LSTM, CNN), blockchain hybrids, and new neuro-symbolic methods. Our meta-analysis shows that hybrid models demonstrate the best detection rates (>99% correctness), but higher operational latency (>100 ms). Despite this, selected optimized machine learning models showed the best overall performance for phishing with a mean correctness of 97.61%. The primary research gaps included bias when keeping datasets (70% of studies utilized outdated benchmark datasets) and the lack of explainability in black-box methods. We make practical recommendations for addressing these gaps, including the need for” real-world” testing environments, lightweight AI-based architectures, and common evaluation metrics to guide evaluation. These findings provide practitioners with an evidence-based framework for selecting the best-supported AI-based cyber defences to outsmart the ever-more-cunning cyber exit strategies.