A founder shipped an entire SaaS built by an AI tool, with zero hand-written code, then posted within days: "guys, i'm under attack."

Strangers were draining his API keys, bypassing the paid subscription, and writing junk to his database, and he could not stop them. The code the tool produced ran fine. The failure was in the decisions around it that he approved without understanding them.

Enrichlead founder on X (@leojr94_), March 2025, reported by Pivot to AI

He shared how he built it, and the attacks started the same week.

There was no sign-in gate, no limit on requests, and the API keys sat in the browser where anyone could read them. All three were textbook gaps, sitting inside choices the tool had proposed and he had approved.

The AI wrote working code, and the choices wrapped around it went unexamined.

A better model would not have prevented any of it, because the gap was never in the code the model produced.

This was not one unlucky founder; the same gap repeats across AI-built apps.

Researchers keep finding the same causes: secrets shipped to the browser, missing sign-in checks, and data anyone could query. The tools generate the code, and judging the plan is left to whoever approves it.

CVE-2025-48757 (Lovable), May 2025; Cloud Security Alliance research notes, 2026

You do not have to write the code, but you do have to understand the plan an AI tool proposes and judge the decisions inside it.

Every attack above traced back to a single choice, about where a key lives, who gets in, or what a stranger may write, that someone approved without being able to judge it.

A working map of how software runs, enough to understand any plan and judge it.

"I brought in the AI so I would not have to understand the code."

Fair, and you still do not have to write it. Understanding a plan is not the same as authoring one, because you are judging a short list of decisions, where the keys live, who gets in, what a stranger may write, not memorizing a language.

The founder who was under attack could not tell that his keys were public or his door was unlocked, so he approved both. A working model could not have saved him, because the choices it returned were his to judge, and he had no way to judge them.

Each gap in the story maps to a move this part teaches.

These four gaps are chapters in this part. The missing sign-in gate and the missing request limit are layers of their own, taught in The Technology Stack.

Every gap in that story is a chapter in this part.

Seven chapters build the map: what the instructions are, where they run, how a request travels, what the language and framework lines mean, how to undo an AI edit, and how a change reaches users. By the end you can read the plan an AI tool proposes and approve each line for a reason you can say out loud.