Generativní jazykové modely a malware: hodnocení rizik, detekce a obranné přístupy

Abstract

This thesis examines the security risks associated with the rapid expansion of Large Language Models (LLMs) and their misuse in the field of malware creation and distribution. The theoretical part portrays the evolution of malicious code from historical file infections to modern "Living off the Land" methods and fileless attacks. The analytical part of the thesis assesses risks associated with AI-assisted phishing and attacks on the integrity of enterprise AI systems based on security reports. This part also proposes strategies for deploying modern AI defensive tools and provides recommendations for the safe use of LLMs in corporate environments. The practical part presents an experimental comparison of current models (GPT-5, Gemini 2.5 Pro, Grok 4, and others) in terms of their susceptibility to producing malicious code and their capacity for obfuscation. Analysis results from the VirusTotal tool demonstrate the effectiveness of LLMs in modifying code to successfully evade static detection mechanisms.