Abstract
Large Language Models (LLMs) are profoundly reshaping the cybersecurity landscape, creating a dual-use dilemma where the same capabilities that enhance defenses also empower sophisticated offensive threats. While security teams leverage LLMs for automated threat detection, static code analysis, and DevSecOps integration, adversaries weaponize them to generate polymorphic malware, zero-day exploits, and highly scalable social engineering payloads. This survey is the first to systematically unify these offensive and defensive perspectives. Drawing on a comprehensive analysis of over 80 academic, industry, and threat intelligence sources, we project the share of LLM-assisted malware to rise from an estimated 2% of all attacks in 2021 to 50% by 2025. We synthesize key countermeasures from platforms like Google Play Protect and Microsoft Security Copilot, and we propose a robust governance roadmap that addresses the technical, ethical, and policy challenges of secure, privacypreserving, and scalable LLM deployment. Our findings position LLMs not merely as tools, but as critical infrastructure-central to both attack and defense-in the emerging AI-powered cybersecurity era. Impact Statement-This work delivers the first comprehensive framework uniting offensive and defensive perspectives on Large Language Models (LLMs) in cybersecurity. By combining technical strategies with governance requirements, it provides actionable guidance for integrating LLMs into secure development, automated threat detection, and policy enforcement. The proposed roadmap enables governments, enterprises, and security teams to deploy LLMs in ways that are scalable, privacy-preserving, and resilient against adversarial misuse. Its synthesis of cross-industry case studies, risk projections, and regulatory trends equips decision-makers with a blueprint for mitigating dual-use risks while maximizing the societal and economic benefits of AI-driven security innovation.