AI in Cybersecurity: How Large Language Models Classify Risk and Opportunity
Alexandru Răzvan Căciulescu, Alin Bărăitaru, Alexandru Radovici · 2025
This study examines the convergence and divergence of large language models (LLMs) in detecting risk and opportunity framings of artificial intelligence (AI) within cybersecurity discourse. Using a structured measurement framework of 18 binary-coded indicators, we prompted seven LLMs to analyze three high-engagement Reddit threads discussing AI in cybersecurity. All models received identical input and were instructed to apply the framework without inference or contextual extrapolation. The analysis shows that while models tend to agree on broadly discussed operational claims, such as AI’s capacity to automate tasks or enable attacks, they diverge on more context-dependent indicators, including institutional reliability, auditability, and human oversight. These findings show that LLMs reflect stabilized discursive patterns but vary in interpreting embedded or contested evaluations. The study contributes to methodological research on LLM-based content analysis and highlights the need for caution when using automated annotation tools in domains involving evaluative and normative judgments.