The article discusses the issue of confidential information leakage when using robotic process automation (RPA) technologies to automate some business processes. This issue is relevant due to the lack of strict security measures in the development of RPA solutions, which can result in the unauthorized dissemination of trade secrets, personal data and other sensitive information. The goal of this work is to develop and test a method for preventing leaks by integrating a system that verifies the data intended to transfer outside the company using small language models into RPA workflows. The PIX Studio development platform was used for the implementation and testing of the proposed method. In the course of our research, we proposed an architecture for a secure RPA solution. We developed a special test set to detect confidential information in heterogeneous corporate data. To evaluate the effectiveness of our solution, we experimentally tested five small language models using precision, recall, and F1-score metrics. The Gemma and Mistral models demonstrated the best results, confirming the feasibility of using small language models for local data analysis without transferring information to external servers. The practical significance of this work lies in the proposal of a ready-made leak control mechanism that can be integrated into RPA platforms. This corresponds with the current trends in intelligent automation and the implementation of DevSecOps (development, security, operations) principles in corporate processes.