Nobody skipped a gate. Every gate had been passed.
The feature shipped on a Thursday. Nine days later a customer's security team found it would apply a credit to an account it had never been asked about. The director of product security pulled the release record expecting to find the gate someone had gone around. He found instead that the sprint had closed forty one changes, thirty of them written with an AI assistant, and the two reviews the lifecycle required had each taken about eleven minutes.
That is the problem this book solves, and it has two halves that arrive together. The first is rate. Assisted and agentic development roughly doubles the volume of change while review capacity stays flat, so the review function absorbs the excess by reviewing less. The second is kind. What ships now contains models, prompts, retrieval, tools, agents, and memory, whose failure modes a reviewer trained on code cannot see in a diff. Adding reviewers does nothing for kind. Adding AI specific review steps makes the rate problem worse. This book names the condition the Dual Impedance Problem and closes it.
Secure by Design in the Age of AI installs a five part operating model. The AI Attack Surface Ledger sees the whole product, one row for every place trust changes hands. The AI Feature Tier Map sorts every feature by what it can reach, change, and cause, so scarce human judgment goes where the consequences are. The Two Lane Assurance Model verifies at the rate the product changes, with a machine lane that blocks on every change and a human lane reserved for what only a person can judge. The Ship Decision Record decides each release on current evidence, with the name of the executive who accepted the residual risk. The Post Release Response Ladder holds the product for its whole life, from the first outside report to retirement.
Ten further instruments support the five, including a maturity baseline, a review capacity metric that gets the budget approved, a component admission standard for models and datasets, a protocol for vulnerabilities that have no code defect, and a single evidence pack that answers the customer, the auditor, the regulator, and the acquirer from one source.
Written for CISOs and the directors and vice presidents of product security who answer to them. Every chapter ends with the board question you must be able to answer, the evidence required, the failure pattern that breaks it, and a thirty day move with a named owner. The final chapters sequence the first ninety days, so your first fully documented release ships in week twelve. No vendor is named. Every executive level claim carries a citation.
Your customers already know what evidence to ask for. This is how you have it ready before they ask.
Volume VII of The Operating Discipline for AI Library.
