Critical infrastructure protection represents one of the greatest challenges of the past decade due to new cyber and physical threats, as well as those generated by climatic, geopolitical, or other forms of instability. This complex of threats can disrupt interconnected and complex systems, leading to cascading failures that can no longer be easily remedied. Therefore, critical infrastructures, which include essential systems aiming at economic activities, public health, and national security, should be analyzed not only from the perspective of reliability, but also from that of their capacity to adapt in case of a major crisis. The study aims to analyze the capability of large language models (LLMs) to prioritize threats to critical infrastructures and to compare these assessments with those made by certified experts in the field. The analysis is based on a set of ten threats and three criteria: the effect on society, economic consequences, and potential casualties. These are ranked through three different LLM models (ChatGPT, Gemini, and Claude), using repeated iterations, while in parallel, the same set of threats is evaluated by a sample of human specialists. Through this approach, it is intended to determine the points where these hierarchies converge, as well as their points of divergence, in order to understand to what extent artificial intelligence can be integrated into decision-support mechanisms designed for resilience and crisis management.