Vulnerability Analysis of Language Models and Development of Compensating Recommendations for the Educational Process

Abstract

This paper is devoted to the urgent problem of ensuring the information security of students in an educational environment using artificial intelligence (AI) technologies. The aim of the study is to develop recommendations for the preparation of an industrial request aimed at minimizing the risks of children receiving unreliable, harmful or inappropriate information, as well as to avoid 'hallucinations' of AI. As part of the work, the analysis of available language models (TryChatGPT, Deepseek, GigaChat) is carried out, key vulnerabilities and potential threats are identified. Based on this analysis, comprehensive recommendations are formulated. The practical part of the work includes an experimental verification of the effectiveness of the proposed recommendations. The results of the study demonstrate a statistically significant decrease in the number of cases of generating false information when applying the developed set of measures. The work can be used to create secure educational platforms and develop methodological manuals on digital hygiene.