What “safe to ship” actually means for software you built with AI
“It feels fine” is not a security statement. Neither is “the model wrote it”. Safe to ship is a short list of checks you can name, re-run after every change, and show to the next person who has to trust the app. Here is that list and the honest edge of it.
1 Safe to ship is a claim about the current build
“We checked it in March” says nothing about the binary running today. A generated app changes on every prompt: a new table, a new route, a new key pasted into a config. If the check does not belong to the build that is live, it is a memory, not a control.
The practical test: can you name the version that was checked, and the findings that were open when it was cut? If not, the claim is unfalsifiable — and unfalsifiable is not safe, it is just unchallenged.
2 The four questions that decide most real incidents
Where does user data live, and what rule decides who may read it? In a hosted backend, every table holding user data is in scope. A rule that reads using (true) is an open door with a nice name on it.
Which secrets can the browser reach? Anything the client can read, a stranger can read. A key that ships in a bundle is already public; rotating it is the fix, not hiding it.
Does the login actually authorize anything? Authentication proves who you are. It does not prove what you may touch. A signed-in user who can read another user's rows has an access-control hole dressed as a feature.
Is the thing real? Generated apps ship stubs — a calculation that returns a constant, a screen that only works on the sample. The demo passing is not the product working.
3 A review produces evidence, not a verdict you stop reading
A useful review names the finding, the file or the table it lives in, and what would close it. It gives you a list you can act on, not a single word that ends the conversation.
Read the findings. A pass that nobody read is the same failure mode as a red flag nobody fixed. The value is the specifics: which rule, which key, which route.
What a review does not give you. No review makes an app provably secure, and none of this is a legal, compliance, or insurance conclusion. It is a check against known failure modes, run at a moment in time, worth re-running when the code moves again.
4 When “good enough” is a legitimate answer
A hobby tool holding nothing personal needs a smaller answer than an app taking payment details or medical notes. The honest question is scale-matched: what would it cost you if this specific data leaked today? Say the number in writing. “Safe to ship” is that number, plus the checks that hold it.
5 The habit that keeps the claim true
Re-run the check when the app changes, and keep the record with the build. That single habit is what turns “I think it is fine” into something you can hand to a customer, an investor, or your own future self — the difference between an opinion and a receipt.
The build review checks these failure modes and returns the findings with the evidence behind them.
Before that, the working sequence is worth reading in full: how to secure a vibe-coded app, the ten holes to check, and what to do when the app is already live and leaking: build rescue.