Artificial Intelligence in Cybersecurity: Techniques, Limitations, and Research Directions

Artificial intelligence (AI) is increasingly used to support cybersecurity tasks such as network
intrusion detection, malware analysis, threat intelligence, and security-operations workflows. At the
same time, AI-enabled systems introduce risks related to model opacity, adversarial manipulation,
privacy, distribution shift, and misuse of generative models. This narrative review organizes the topic
into four connected areas: AI-based detection, explainable AI (XAI), adaptive threat intelligence, and
generative AI and large language models (LLMs). Across these areas, the central engineering
challenge is not only to detect malicious activity, but to produce reliable, interpretable, and operationally
useful outputs under changing conditions. The review highlights recurring limitations in evaluation
realism, explainability, adaptation to evolving threats, and human oversight. It proposes a practical
research agenda focused on reproducible benchmarking, robustness testing, human-in-the-loop
response, privacy-aware data practices, and transparent reporting. Because this manuscript is a
narrative review rather than a systematic review, it does not claim exhaustive coverage of the literature.
A future version should document a reproducible search and screening protocol and verify all
bibliographic records before journal submission