The rapid deployment of Large Language Models (LLMs) across enterprise and consumer applications has introduced a novel class of security vulnerabilities that challenge conventional cybersecurity paradigms. This paper examines the emerging AI security landscape through three interconnected lenses: the transplantation of the Red Team / Blue Team methodology from classical information security into the domain of AI; the taxonomy and evolution of jailbreaking techniques used to circumvent LLM safety alignment; and prompt injection—now ranked as the #1 threat in the OWASP Top 10 for LLM Applications (2025)—as the dominant unsolved attack vector threatening LLM-integrated systems. We further analyze the intersection of traditional vulnerability classes such as Insecure Direct Object Reference (IDOR) with AI-augmented offensive tooling, the threat of knowledge-base poisoning in Retrieval-Augmented Generation (RAG) systems, and the current state of AI-as-firewall defensive architectures. Our analysis synthesizes practitioner observations and peer-reviewed findings to argue that AI security constitutes an unending adversarial cycle—one that demands continuous, layered, defense-in-depth strategies rather than static, checklist-driven postures.