SECURITY OF LARGE LANGUAGE MODELS: RISKS, THREATS, AND SECURITY APPROACHES

The article provides a comprehensive analysis of current security challenges related to Large Language Models (LLMs), which have become a key element of digital transformation across multiple sectors. It examines typical threats arising both from targeted attacks on models and from their malicious use in cybercrime. The main risk vectors are identified, including prompt injection - embedding hidden instructions in user queries to alter model logic, and jailbreaking - crafting prompts that bypass built-in restrictions and trigger undesirable behavior. Special attention is given to the risks of confidential data leakage from training datasets, generation of vulnerable or malicious code that can enter production environments, and the dissemination of disinformation, including multimedia deepfakes. Based on the analysis, a conceptual LLM security model is proposed, combining technical, architectural, and regulatory elements of protection. Particular emphasis is placed on assessing and applying mechanisms such as AI firewalls - intermediary systems that filter model inputs and outputs; built-in security modules integrated into model architectures; and guardrails - restrictions on outputs without altering parameters. Additionally, watermarking methods for synthetic content identification and AI-generated content detection tools are considered. Regulatory measures are highlighted as an essential component to establish usage frameworks for powerful models and mitigate misuse risks. It is concluded that traditional cybersecurity measures, focused on static or signature-based detection, are insufficient for generative systems operating in dynamic natural language environments. Enhancing security requires multilayered strategies covering all stages of the LLM lifecycle - from design and training to deployment and regulatory oversight.