Application of Large Language Models for Assessing Parameters and Possible Scenarios of Cyberattacks on Information and Communication Systems

Abstract

This paper explores the use of large language models (LLMs) to evaluate parameters and identify potential hostile penetration scenarios in corporate networks, considering logical and probabilistic relationships between network nodes. The developed methodology is based on analyzing the network structure, which includes components such as the Firewall, Mail Server, Web Server, administrator and client workstations, application server, and database server. The probabilities of transitions between these nodes during adversarial attacks are determined using a swarm of virtual experts and two sets of prompts aimed at different LLMs. Among the results obtained through the swarm approach are average transition probabilities, which enable modeling the most likely attack paths from both external and internal network origins. Based on logical-probabilistic analysis, penetration scenarios are ranked according to probabilities, execution time, and resource minimization required by attackers. The proposed methodology facilitates rapid response to threats and ensures an adequate level of cybersecurity by focusing on the most probable and dangerous attack scenarios.