Title: Generative AI in Cybersecurity: From Automation to Adversaries

Abstract: Generative AI is becoming an important factor in cybersecurity. It can support defensive tasks, accelerate offensive workflows, and create new attack surfaces of its own. This tutorial examines the capabilities, limitations, and security implications of modern generative AI systems from a technical perspective.
We begin with the foundations most relevant to cybersecurity researchers, including large language model architectures, self-supervised pre-training, alignment, prompting, retrieval-augmented generation (RAG), and tool-using agent workflows. We then present two case studies chosen to move beyond demonstrations and to expose the limitations that appear in realistic security settings. The first studies an LLM-based agent used in network-security laboratory tasks such as reconnaissance, evidence collection, report generation, and man-in-the-middle experimentation. It shows where automation is effective, where it fails, and which failure modes recur in practice. The second examines the use of LLMs together with formal verification tools to translate informal secure-messaging protocols into formal models, help formulate security queries, and interpret proofs and counterexamples.
The two case studies lead to a common conclusion. Generative AI can accelerate parts of cybersecurity workflows, but correctness, assurance, and sound security reasoning still depend on external validation, formal methods, and human oversight. They also point to a set of research problems, including trustworthy tool use, evidence-grounded automation, LLM-assisted formal modeling, and the security of generative AI systems themselves. The tutorial closes with a discussion of attack surfaces, recurring failure modes, and research directions spanning penetration testing, formal security analysis, and adversarial attacks on generative models.