Machine Learning and Watermarking for Accurate Detection of AI-Generated Phishing Emails

Large Language Models offer transformative capabilities but also introduce growing cybersecurity risks, particularly through their use in generating realistic phishing emails. Detecting such content is critical; however, existing methods can be resource-intensive and slow to adapt. In this research, we present a dual-layered detection framework that combines supervised learning for accurate classification with unsupervised techniques to uncover emerging threats. In controlled testing environments, our approach demonstrates strong performance. Recognising that human users are often the weakest link in information security systems, we examine historical deception patterns and psychological principles commonly exploited in phishing attacks. We also explore watermarking as a complementary method for tracing AI-generated content. Together, these strategies offer a scalable, adaptive defence against increasingly sophisticated phishing attacks driven by Large Language Models.