ABSTRACT. The growing use of artificial intelligence in cybersecurity has increased the necessity of transparency, interpretability, and accountability. The Explainable AI (XAI) provides a way to mitigate those issues, but it has not been fully developed or implemented in the security fields yet. The paper presents a systematic comparative synthesis and critical analysis of 50 articles on malware detection, intrusion detection systems, hybrid XAI models, and regulatory or ethical frameworks. These studies are assessed using a methodological framework on the aspects of technical rigor, human-centered design, and governance alignment. The findings indicate that there is a strong dependence on post-hoc methods like SHAP and LIME, which, despite their flexibility, do not necessarily have fidelity, scalability, and adversarial robustness. Intrinsic and hybrid models are more promising as far as interpretability can be embedded without reducing accuracy, but can only be validated on fixed benchmark datasets, which is limiting to real-world applicability. A significant gap in human-centered assessment is also indicated by the review: not many studies determine the effect of explanations on the trust calibration, cognitive load, or operational performance of analysts. Regulatory and ethical aspects are also largely under-researched, and the majority of implementations do not respond to the needs of compliance with standards like the GDPR and EU AI Act. Taken together, these results indicate the disjointed nature of XAI in the field of cybersecurity and the dire necessity of combined frameworks that can synchronize technical innovation with human usability and compliance with governance.