ANALYSIS OF NEURAL NETWORK PATTERNS IN THE GPT-3.5 LANGUAGE MODEL DURING ITS USE BY MALICIOUS ACTORS FOR PHISHING ATTACKS

Dmitrieva Viktoriya Alexandrovna, Shatov Igor Alekseevich, Batyanova Darya Denisovna, Fedorov Vadim Valeryevich

This article is dedicated to the investigation of the capabilities of a neural network based on the OpenAI GPT-3.5 language model in the context of its potential use by malicious actors for generating phishing emails. Particular attention is given to identifying patterns and specific characteristics that emerge when this model is employed in phishing attacks. The aim of the study is to analyze the structure and features of phishing messages created using the neural network, as well as to determine approaches for imitating typical phishing templates in both formal (business) and informal (friendly) styles. The research involved 43 experiments, each consisting of a series of messages generated by the neural network. The analysis focused on criteria such as the reasons for writing the email, subject lines, sender and recipient identities, and communication style. The research methodology included a detailed examination of the text, formatting practices, and the model’s ability to adapt its writing style based on predefined parameters. The results demonstrated that the neural network is capable of effectively generating emails in a formal business style; however, in the informal style, messages often lose naturalness, which reduces their credibility. In conclusion, the article highlights the importance of continued research into such technologies—not only in terms of their advancement but also in the broader context of information security.

Keywords:phishing, neural networks, OpenAI GPT-3.5, natural language processing, social engineering