THE DUAL ROLE OF LARGE LANGUAGE MODELS IN CYBERSECURITY: THREATS, DEFENSES, AND RESEARCH DIRECTIONS

Large language models (LLMs), have come in from three angles into the cybersecurity world. They're a new class of system that has to be secured, they are a capability that reduces the cost of offense, and they are a tool that enhances defense. Current reviews consider either of these directions only, making it hard for researchers and practitioners to glimpse the interplay of the three. This review collates research published from 2023–2026 in all three of the following, presented in the dual-use approach. We provide a three-layer taxonomy of attacks on LLM-based systems, which include attacks at the model layer (jailbreak, poisoning, extraction), the system layer (direct, indirect, and multimodal prompt injection, retrieval poisoning), and the agentic layer (memory poisoning, tool abuse, protocol exploits, and multimodal agent propagation). Then, we discuss the use of LLMs for attacking (phishing, social engineering, malware generation, autonomous attacks) and protecting (threat intelligence, intrusion detection, vulnerability detection and repair, security operations). We list the key defense families and the standards of measurement for them. There is a common theme in our synthesis: the reported success rates of similar threat classes vary from less than 10% to more than 87% in different benchmarks and experimental designs; threats are rarely tested against adaptive threats; and the security to utility trade-off is rarely quantified. We conclude with eight research directions: Multilingual Guardrail Robustness, Longitudinal Security Measurement and Human Factors in AI-assisted Security Operations.

Hali tarjima qilinmagan