Year: 2025
Type: dissertation
Abstract: Multi-agent systems composed of language agents have shown promising results in various problem domains. Language agents leverage the capabilities of Large Language Models, such as in-context learning, reasoning, planning, and decision-making, addressing previous challenges in agent research and development. Both Large Language Models and language agents have demonstrated effectiveness in cybersecurity tasks, including threat detection and incident response. This work explores the application of multi-agent systems to cybersecurity by using the automation of security incident analysis as a case study. An architecture for building multi-agent systems by composing small, specialized agents is proposed. The architecture is designed based on a literature review of agents in artificial intelligence and objectives distilled from discussions with experts. These objectives highlight many practical challenges in developing language agents and multi-agent systems to solve complex problems. Using the architecture, a multi-agent system for summarizing processes is implemented and then evaluated using real security incidents, showcasing the architecture's effectiveness. To further demonstrate the applicability of the architecture to other problem domains, an existing multi-agent system is redesigned to use the architecture. The results indicate that the architecture is suitable for building multi-agent systems in various domains, including cybersecurity.
Source: Aaltodoc (Aalto University)
Author: Palojärvi, Oskari
Institution: Aalto University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7076971186
