Експлуатація вразливостей штучного інтелекту та способи їх запобігання

Abstract

The growing integration of AI into critical infrastructure introduces new cybersecurity threats targeting large language models (LLMs) and agentic AI systems. Threat actors now embed LLMs directly into malware for real-time code generation and self-obfuscation, while autonomous AI agents can perform up to 90% of intrusion operations with minimal human involvement. This paper classifies AI vulnerabilities into model-level, data-level, prompt-based and orchestration-layer categories, examines social engineering techniques for bypassing model safeguards, and proposes mitigation measures including secure AI development lifecycles, risk management frameworks, monitoring with killswitch mechanisms, and defensive AI-driven threat detection