Efektivita a komparace ofenzivních a defenzivních postupů velkých jazykových modelů v oblasti kybernetické a fyzické bezpečnosti

Abstract

This thesis examines the effectiveness and comparison of offensive and defensive procedures formulated by large language models (LLMs) in cybersecurity and physical security. The theo-retical part analyzes LLM architecture, their offensive and defensive applications, and situates the topic within a regulatory and economic context through PESTLE analysis. The practical part experimentally compares five models (GPT-5.2, Claude Sonnet 4.6, DeepSeek-V3.2, Grok 4.1, Mistral Large 3) in generating security procedures and evaluates the quality and practical applicability of their outputs.