A panel discussion on "Scientific, Technical, and Innovative Support for National Security" was held as part of the business program of the "National Security. Belarus 2026" exhibition. Experts from the United Institute of Informatics Problems of the National Academy of Sciences of Belarus participated in the event.

Andrei Bondolovsky, Head of the Speech Recognition and Synthesis Laboratory, presented a paper on "Creating Large Language Models: An Important Factor in National Information Security." The expert identified four key dimensions of information security associated with the use of large language models:

The model processes request data: the transfer of data abroad (correspondence, code, analytics, and personal data) creates risks of information leakage;
The model's infrastructure: dependence on external services makes the infrastructure vulnerable: the shutdown or failure of such services can become a tool of pressure;
The formation of worldview: language models shape attitudes and assessments in responses, which can serve as a tool of soft power and information influence, influencing public opinion;
A model may contain "bookmarks"—hidden vulnerabilities or misconfigurations that are often difficult to detect without access to the model's internal weights and data. In this case, trust in the system becomes crucial.

Vasily Kovalev, a leading researcher at the Biomedical Image Analysis Laboratory, presented a report titled "Security Issues in Neural Network Solutions: New Challenges in the AI ​​Era." He emphasized that the threat of leakage and misuse of personal data is gradually and silently permeating our lives, and recognizing this problem is the first step toward solving it. The expert noted that artificial intelligence and neural network solutions present numerous vulnerabilities related to attacks on models. Protecting AI solutions is a complex task requiring ongoing effort. For over 10 years, the Joint Institute for Informatics and Problems of Information Processing (UIIP) of the National Academy of Sciences of Belarus has been developing neural network security systems, including model "immunization" processes—testing and training them against potential attacks—to reduce the risk of misleading them. Vasily Alekseevich added that it is crucial not only to develop secure systems but also to implement procedures for continuous testing and auditing of their security to ensure a high level of resilience to external threats.

Discussions held at the "National Security. Belarus 2026" exhibition highlighted the pressing relevance of information security issues in today's environment. Developing reliable data protection technologies, creating resilient systems, and recognizing potential risks are key tasks for science and the state in ensuring national security.