An agent-based solution to prevent data leakage in-use state

It is very important for every institution, especially public institutions, to ensure the confidentiality of the data they use, especially in internal correspondence and communication, for the healthy functioning of the institution and the sustainability of competition between institutions. The issue of ensuring data confidentiality has been the subject of many scientific studies and as a result of these studies, various solution proposals have been proposed depending on the status and location of the data. With the increasing importance of the concept of cyber security, all solutions found and proposed need to be updated periodically, become flexible and pass through the scale of applicability both locally and globally. Many solution proposals currently offered have come to the point of losing their validity after the widespread use of the concept of artificial intelligence and the emergence of large language models. Another problem is that the solutions in the literature are not suitable for the Turkish language. In this study, first of all, the dictionary-based structure, which is one of the data leakage prevention methods, was localized and implemented with Turkish language support. Using an agent software installed on in-house computers, the employee's correspondence was instantly checked with the dictionary system, and real-time reporting was made using a real-time cloud database, and it was aimed to ensure that precautions were taken when a leak occurred. Secondly, correspondences collected in certain blocks were sent to the large language model in order to test the existence of personal data with artificial intelligence and an attempt was made to detect the presence of sensitive data in the text. The role structure of the large language model was used to ensure that the study could be organized according to different sectors and thus flexibility was added to the application. The personal data violation detection status of the large language model and the prohibited word violation detection were reported to the system administrator.