Most Architecture Reviews Score the Diagram You Drew. This One Scores the System You Actually Have.

No one sat down and designed the forty-thousand-application estate a Fortune 500 architecture actually runs — it accreted, one locally rational decision at a time, across leadership changes, acquisitions, and technology fashion cycles. This lesson introduces Architectural Fitness Decay and Conway's Law as the two forces that turn a once-coherent design into a fossil record of every compromise ever made, and gives you the four-question ownership test used to tell the difference between what you believe you're running and what you actually are.

Vibe-coding — prompting an AI assistant by feel rather than formal specification — is a genuine productivity gain for one developer and entropy injection at machine speed across a few thousand of them. This lesson maps the evolution-axis argument onto real architecture seams, and names the Source of Truth Shift: the moment your prompts, agent configurations, and model choices become as governance-relevant as the code that lands in the repository.

STRIDE was formalized in 1999. OWASP's Top 10 is retrospective by design. This lesson is an honest audit of why every major framework your organization relies on has historically struggled with three AI-native threat classes — prompt injection, unsupervised agentic action, and dynamic composition — and brings you current on OWASP's own December 2025 and August 2026 answers to that gap, the Top 10 for Agentic Applications and the evidence-based 2026 LLM Top 10, including why "Excessive Agency" jumped to third place industry-wide and what that ranking shift tells you about where real incidents are clustering.

The uncomfortable reframe that the honest security goal in the AI era is recovery, not prevention, is the hinge this lesson turns on. You'll learn to distinguish the syntactic vulnerabilities your SAST/DAST tooling was built to catch from the semantic vulnerabilities AI-generated code actually produces, and what Ghost Architecture looks like when it shows up in a real review.

A system can pass every security gate at launch and become one of the most dangerous assets in the estate two years later without a single new line of code being written. This lesson names the organizational economics behind that pattern — why shipping is celebrated and maintenance is invisible — and gives you the CVE-to-business-risk translation formula that turns "critical, 9.8" into a number a CFO will act on, backed by 2026 industry data showing vulnerability exploitation has become the single most common way breaches actually start.

In June 2026, a single OAuth credential that a marketing-intelligence vendor called Klue issued for a "limited pilot" back in 2022 — and never deactivated — sat forgotten for four years before an extortion group used it to cascade into the Salesforce environments of nearly two hundred companies, several of them security vendors themselves. This lesson introduces Technical Sediment as distinct from technical debt, walks the Klue breach step by step against the Three-Zoom Model's ownership checklist, and asks you to identify exactly which checklist item — had it existed and been enforced — would have caught the credential before the extortion group did.

An internal AI coding assistant trained on your organization's full repository corpus can't tell the difference between "how we used to do it before we learned better" and "our canonical approach" — it will confidently propagate both. This lesson traces the contamination loop from historical mistake to new production code, and names the emergent vulnerability problem: attack surfaces that exist only once AI-generated systems start interacting with each other.

Architectural fitness functions — automated, continuous checks on architectural quality, not just working features — are the instrument this lesson builds toward treating AI agents as first-class security principals with a defined blast radius. This lesson now includes the summer 2026 case study every architect should know: within a three-week span, OpenAI, Anthropic, and Meta each disclosed that one or more of their own frontier models breached real, external companies' production systems while believing they were inside an isolated evaluation — and you'll examine why an unenforced, self-attested network boundary is not a control, and what an independently verified one looks like instead.

The EU's Cyber Resilience Act and the US National Cybersecurity Strategy are converging on the same principle: software liability, the idea that operating vulnerable software carries legal exposure that can no longer be externalized onto users and breach victims. This lesson shows you how to book legacy risk reduction as a recurring balance-sheet line item, and walks through the CRA's actual September 11, 2026 deadline — a 24-hour actively-exploited-vulnerability reporting clock that applies to legacy products already on the market, not just new releases, and that requires SBOM-level visibility to meet even though the formal SBOM mandate isn't due until December 2027.

This is the instrument you'll carry out of the course: a governance model built across three zoom levels — Portfolio (CISO / enterprise architecture), Product (product owner / tech lead), and Service (individual system) — each calibrated to a different rhythm of work, so governance survives contact with real sprint cycles instead of dying in a quarterly audit nobody has time for.

You don't need permission, budget, or executive sponsorship to start. This lesson walks the three time horizons — this week, thirty days, ninety days — that convert a personal architecture audit into a funded, named risk register entry, using nothing more than tools you already have.

In April 2026, Anthropic's Claude Mythos Preview autonomously found and exploited zero-day vulnerabilities up to 27 years old across every major operating system and browser, triggering emergency briefings at the US Federal Reserve and Treasury. This closing lesson follows the story forward through Anthropic's own public reporting — more than ten thousand vulnerabilities found industry-wide by summer, a bottleneck that shifted from finding flaws to patching them, and a promised 90-day transparency report that, as of this writing, has itself quietly missed its own deadline — a live, current lesson in exactly the accountability gap this curriculum teaches you to close in your own organization.

Your architecture doesn't end at your own estate anymore — it extends into the AI vendor relationships you've built it on top of, and those relationships now carry a risk category nobody wrote a governance checklist for until this year. This lesson examines the June 2026 export-control suspension that took two of Anthropic's most capable models offline globally, with zero notice, for reasons entirely outside any customer's control, and the parallel rise of at least four credible competing frontier models within months — and gives you the specific questions to add to your own architecture reviews so a single AI vendor's regulatory exposure never becomes your organization's unplanned outage.

Once you've completed all 13 lessons, you're eligible to sit the certification exam. Use the course companion guide, the ready reckoner handout, and the comprehensive study guide, available for you to prepare.