Anatomy of AI Systems and Their Attack Surfaces

As artificial intelligence (AI) systems move from research labs into production environments, they introduce fundamentally different architectures and operational behaviors that redefine what it means to secure software. This chapter explores the structural distinctions between predictive, generative, and agentic AI, explaining how their internal mechanisms shape risk exposure, attack surfaces, and defensive priorities. From supervised models optimized for speed and accuracy to large language models capable of generating text and executing tools, each architecture comes with unique adversarial challenges—ranging from input manipulation to autonomous decision hijacking. The chapter also walks through the AI development lifecycle, from data ingestion to post-deployment monitoring, revealing how vulnerabilities can be introduced at each phase. It emphasizes the importance of recognizing adversarial entry points—not just in APIs or user prompts, but in training data, model fine-tuning, and third-party integrations. Readers gain practical insight into how attacks like model poisoning, prompt injection, and supply chain compromise exploit architectural and process-level blind spots. Crucially, this chapter introduces the concept of security debt—accumulated risk resulting from rushed development, lack of robustness testing, or weak observability. It shows how ignoring security early in the lifecycle compounds risk downstream, often surfacing only during high-stakes deployment. By combining architectural literacy with lifecycle threat awareness, this chapter equips cybersecurity professionals to proactively assess and defend AI systems before adversaries do.