5 vibe coding mistakes that turn a demo into an incident
In February 2025, Andrej Karpathy — the AI researcher who helped build OpenAI and then ran Autopilot at Tesla — tweeted a sentence that became a movement: “There’s a new kind of coding I call “vibe coding”, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.”
Two years on, vibe coding has shipped real products, real hacks and a fair number of real databases dropped in production. The failure mode is not mysterious. The speed is real; the missing part is everything that used to happen between “it works on my screen” and “it works for everyone”. Here are the five mistakes we keep untangling in vibe‑coded projects — and they are all fixable without giving up the speed.
In short
- Shipping unreviewed code is the root error; everything else grows from it.
- The numbers back the concern: 65% of 1,400 scanned vibe‑coded production apps had security issues, 58% had at least one critical vulnerability (Escape.tech).
- Secrets in prompts — 400+ exposed credentials found in that one scan. Your API keys are not prompt context.
- No tests, no gate — “it worked in the chat” is not a deployment criterion.
- No version control discipline — vibes without git mean nobody can tell what changed.
- Wrong layer for vibing — authentication and payments are not the place to embrace exponentials.
Mistake 1: shipping the demo
Vibe coding is a prototyping superpower. The trouble starts when Friday’s demo quietly becomes Tuesday’s production system — same code, same shortcuts, now handling real users and real data.
The best data point so far: API security firm Escape.tech scanned over 1,400 vibe‑coded production applications and found that 65% had security issues and 58% contained at least one critical vulnerability, including over 400 exposed secrets and 175 instances of exposed personal data such as bank account details. These are not hobby projects abandoned on a laptop — they reached production like that.
The fix costs one line of process: nothing goes to production without a human who reads it. Not “glanced at”, read — with an eye on auth checks, input validation and where data flows. The AI writes faster than any reviewer can read, so the review is the bottleneck. That is fine. The review is the job now.
Mistake 2: telling the AI your secrets
The most avoidable finding in that scan: 400+ exposed secrets. It happens exactly as you’d guess — a developer pastes an .env file, a connection string or an API key into a chat window because the assistant needs context to “fix the auth”.
Treat every prompt as a message to a third‑party service that may be logged, may be retained and may surface in someone else’s context window tomorrow. The rule is boring and absolute: credentials live in environment variables and secret managers; prompts get placeholders. When an AI session needs to debug something key‑shaped, the human rotates the key afterwards regardless. Rotation is cheaper than incidents.
Mistake 3: “it worked in the chat” as a test strategy
Vibe‑coded features fail in a specific, sneaky way: they handle the demo path perfectly and fall apart on the paths nobody demoed. Empty inputs. Concurrent users. A second language in the content. A timezone.
The antidote is unglamorous: every generated change gets a test that fails without it, and a CI gate that refuses to deploy red. If writing a test feels slower than accepting the change, that feeling is precisely the discipline paying for itself later. Tests are also the honest interface between you and the AI: “make the failing test pass” is a far better prompt than “improve the code”.
Mistake 4: no git, no history, no idea what changed
“…and forget that the code even exists” was a joke about focus, but some teams read it as workflow advice. The result shows up in our inbox as a question we can’t answer for them: “the site worked on Thursday, something changed, can you help?” — with no commits, no diff, no history.
The AI doesn’t need you to ignore version control; it works better with it. Small commits after every working state, meaningful messages, branches for anything risky. Then when an agent goes sideways — and one eventually will — you revert instead of archaeologizing. Git is the undo button that makes aggressive experimentation safe. Skipping it doesn’t make you faster; it makes you braver than you can afford.
Mistake 5: vibing the wrong layer
Not all code is equally safe to generate without deep review. Vibing a marketing landing page is low‑stakes: worst case is an ugly page. Vibing authentication, payment logic, data deletion or anything handling personal data is a different sport: worst case is the Escape.tech report.
Our split is simple. Vibe freely: UI, copy pages, prototypes, internal tools, one‑off scripts, migrations with dry runs. Never vibe blind: auth, payments, PII handling, infrastructure with destructive permissions. Prototype those by all means — then implement them deliberately, review hard and keep a human accountable for each line.
The grown‑up version exists
None of this says stop. It says: the harness around the AI is the product. Karpathy himself closed the loop in February 2026, looking back at his own term: “It was good fun and it almost worked.” Almost. The teams that made it work kept the leverage and added the oversight — he now calls the disciplined version agentic engineering, and it’s what we practice with our AI agent development service.
If you want the deep reason multi‑agent setups amplify these mistakes instead of fixing them, that’s the follow‑up: why vibe coding fails — too many agents.
Summary
Vibe coding earned its place — as long as the review, the secrets policy, the tests and the git history arrive with it. The 65% and 58% figures are not an argument against AI‑written code; they are a price list for skipping verification. Pay the process tax, keep the speed.
If you’d rather have that harness built in from day one, that’s what we do — see AI agent development and our approach to web development, or start from the full service list.
Yuri Osipov
Founder, SEO & analytics
Runs Take the Space. Specialises in technical SEO, measurable analytics setups and AI search visibility.