Cyber Defence, Agentic AI, and the Sociotechnical Paradigm: A Survey of AI-Driven Cybersecurity
Abstract
Artificial Intelligence (AI) and Large Language Models (LLMs) are reshaping cybersecurity, moving it from signature-based reactive security to agentic, self-evolving, and proactive operations. However, the fast deployment of autonomous agents entails significant security issues, physical-layer attacks in cyber-physical systems, alignment problems, and advanced social engineering risks jeopardizing the stability of the modern sociotechnical infrastructures. This survey paper, which summarizes and analyses the literature, examines the duality of AI in cybersecurity, exploring its implications for strengthening digital defences and increasing the threat landscape for adversaries. We perform a thorough review of 36 cutting-edge research articles, which are organized into six important areas: self-evolving defensive paradigms, LLM-based intelligent agents, reinforcement learning for autonomous decision-making, security of cyber-physical systems, human-machine co-teaming, and agent alignment. Our synthesis shows that multi-agent collaborations greatly shorten incident response times and automate discovery of vulnerabilities but are also highly vulnerable to prompt injection, training data poisoning, and physical-layer spoofing. Moreover, we identify a shift from model patches to model alignment and constitutional safety at runtime as a prerequisite for securing these systems. We believe that resilient digital defence must also be considered a sociotechnical problem, as there must be agents, autonomous systems, that are governed by other agents, aligned and constitutionally bound, and are able to collaborate, rather than operate completely separately, with human operators.