Year: 2025
Type: dissertation
Abstract: Due to recent advancements in natural language processing, Large Language models have emerged as powerful tools that enable the efficient solving of complex natural language processing tasks. These tools have gained interest in various industries for their ability to automate a wide range of time-consuming tasks, usually performed by humans. LLM Agents are advanced AI systems that expand beyond the conventional text processing of LLMs and allow sequential reasoning through the use of external tools that enable the solving of even more complex tasks. These systems are inherently extremely complex, making cybersecurity a significant challenge. The aim of this thesis is to study the application and cybersecurity of LLM Agents. As LLM agents are a relatively new technology, there is no comprehensive framework for their secure implementation. To find the different applications for LLMs, we first research various industries interested in applying the technologies. We conducted a search for papers discussing LLMs, and we classified them by various subject areas to identify which industries are using these technologies. After this, we discuss the different applications in the identified fields. Finally, we discuss the major cybersecurity risks of LLMs and investigate the defense strategies proposed against these risks. In this thesis, we identify Medicine and Engineering as the fields with the most papers published on LLMs. Other fields that are showing interest include Business and Education. Prompt injection, data poisoning, and model extraction are discussed to cover cybersecurity risks from different aspects of these systems. Promising defense strategies for each of them are found. Multi-agent systems are identified as the most interesting defense strategy for further research to assess.
Source: Aaltodoc (Aalto University)
Author: Kalliomäki, Aleksi
Institution: Aalto University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7020179282
