Vyatkin Artem A.

The article discusses three main types of prompt injections: direct, indirect, and those aimed at generating malicious instructions. In the context of LLM application in the public corporate sector, the authors propose a systematic approach to their analysis. The study examines the risks associated with employing LLMs to support electronic document management, prepare official documents, and automate the processing of citizens’ requests. The results obtained indicate the need to develop specialized protection methods and regulations for the use of LLM. They also highlight the need to introduce mechanisms for validating model conclusions and to limit the powers of agent systems.