\cleardoublepage \part{Security} \setbookpart{Security}
Part IV: Security
AI systems inherit every weakness of the software and infrastructure they run on and add their own: their behaviour can be steered through the data they read, their components arrive as opaque third-party artefacts, their agents hold credentials, and adversaries have the same technology. Part IV therefore treats security for AI as an engineering discipline with its own threat model, controls, tests and evidence.
Chapter 16 extends security architecture to the new AI attack surface. Chapter 17 is a handbook for threat-modelling, hardening, red-teaming, monitoring and incident response, and for the security evidence European regulation expects. Chapter 18 treats AI as the source of threat: adversaries who turn published patches into working exploits within days, run intrusions at machine speed and impersonate executives. Chapter 19 contains the organisation's own agents when they exceed their mandate, without depending on the agent's intentions.
The treatment is strictly defensive and provides no attack tooling. Following Architecture as Code, controls are policy and evidence comes from pipelines. How security evidence feeds risk management and audit is covered in AI Governance (Chapters 8 and 11).
What you will learn in this part:
- Controls against prompt injection, poisoning, model extraction and supply-chain compromise
- How to threat-model, harden, red-team, monitor and respond to incidents, and produce evidence
- How to defend as the window between disclosure and exploitation collapses
- How to contain the organisation's own agents without relying on the agent to behave