Vibe coding is not bad in itself. It is a fast way to build prototypes, demos and internal tools, and it turns risky when an app holds real users' data and nobody has checked who can read what. Andrej Karpathy, who named it in February 2025, called it 'not too bad for throwaway weekend projects'. The production evidence points at access control: an NVD record describes insufficient row-level security in apps one AI builder generated, and Veracode reported that 45 percent of AI-written code samples it tested failed security tests. Test access rules, keys and database policies before launch.
The short answer
Vibe coding is not bad. It is the wrong tool for one job: putting an app nobody has read in front of real users. For a prototype, an investor demo, a landing page or an internal tool, prompting an AI builder and accepting what it writes is a fast and sensible way to work. Once the app stores other people's data, takes payments or has to last, someone needs to read and test the parts a demo never touches: who can read which records, where the keys live and what the database lets a stranger do.
The evidence for that line is specific. A record in the US National Vulnerability Database describes insufficient database rules in apps generated by one AI builder. Veracode reported that 45 percent of the AI-written code samples it tested failed security tests. And most developers in Stack Overflow's 2025 survey said vibe coding is not part of their professional work. The sections below give the sources and what to do about them.
Where the term comes from
Andrej Karpathy named vibe coding in a post on X on February 2, 2025. He described a way of building where you "forget that the code even exists". He accepted every change the AI made, stopped reading the diffs and pasted error messages straight back in, which usually fixed them. He also wrote that the code grew past what he could follow without a long read.
His own verdict on where it fits was modest. He called it "not too bad for throwaway weekend projects". Our answer, nearly two years on, is close to his: the method is fine when the stakes are low and the code is disposable.
Today the term is used more loosely. People say vibe coding for almost any app built mostly by prompting a tool such as Lovable, Bolt, v0, Replit or Cursor, whether or not anyone reads the output. Stack Overflow's 2025 Developer Survey used a broad definition, generating software from LLM prompts. Of the 26,564 people who answered, 72 percent said it is not part of their professional development work, and about 5 percent more said so emphatically.
Where vibe coding is fine and where it needs an engineer
The deciding question is what happens if the code is wrong. If the answer is a broken demo, vibe code it. If the answer is a stranger reading your customers' records, have an engineer review it first.
| What you are building | Vibe code it? | Why |
|---|---|---|
| Clickable prototype or investor demo with sample data | Yes | There is no real data to leak, and speed matters more than structure |
| Landing page or waitlist form | Yes, then check one table | The email list is real data, so confirm a stranger cannot read it |
| One-off script or data cleanup | Yes | Run it on a copy of the data first |
| Internal tool for a few colleagues | Usually | Keep it behind company sign-in, and review it before it touches customer records |
| App where users sign in and see their own records | Needs an engineer's review | Access control and database policies decide whether one user can read another's data |
| Anything that takes payments | Needs an engineer | Payment status must come from the payment provider's server, never from the browser |
| Health, financial or children's data | Needs an engineer and a security review | Laws on that data add duties on top of the security basics |
| A product you plan to extend for years | Needs an engineer | Code nobody has read is slow and risky to change |
The table reflects what we see when we read prototype code.
What goes wrong when vibe-coded apps reach production
The failures cluster in four places: database rules, keys in the browser, the generated code itself and long-term upkeep. None of them shows up in a demo, because a demo uses one account and the happy path.
Database rules that let strangers in
When the app talks straight to its database from the browser, the database rules are the security. Lovable's security docs say it generates frontend code that runs in the browser, that row-level security policies control which users can read or change data, and that misconfigured policies are a common cause of data leaks. Supabase's row-level security docs add the detail: a table in an exposed schema without row-level security is readable and writable by any role with a grant on it, adding policies does not remove those grants, views bypass row-level security by default, and the service role bypasses it too, so its key must stay on the server.
A public example is CVE-2025-48757. The National Vulnerability Database record, published May 30, 2025, says an insufficient row-level security policy in Lovable through April 15, 2025 allowed remote unauthenticated attackers to read or write arbitrary database tables of generated sites. It shows a CVSS 3.1 base score of 9.3 out of 10, assigned by MITRE as the reporting authority; NVD itself has not scored it. The record also notes that Lovable disputes it, because each customer of the platform accepts responsibility for protecting their own app's data.
The researchers who reported it published their own write-up on May 29, 2025. Their script visited the homepages of 1,645 projects from Lovable's showcase of launched apps and reported 303 endpoints across 170 of them, about 10.3 percent, with inadequate row-level security settings. They also report inserting a record marked as paid on one site without going through its payment step. Read it as one side of a dispute: its author works at Replit, which offers its own AI app builder, and Lovable contests the record.
Keys and secrets in the browser
Anything shipped to the browser can be read by anyone who opens the developer tools. Lovable's docs say it plainly: "API keys and other secrets cannot be stored safely in client-side code." The researchers' write-up lists anonymized endpoints that exposed this kind of data, with names such as get-google-maps-token and get_gemini_api_key. Keep secrets in server-side functions, and rotate any key that ever reached the browser or the repository.
The generated code itself
Veracode's 2025 GenAI Code Security Report tested over 100 large language models on tasks in Java, Python, C# and JavaScript. Its blog post of July 30, 2025 reports that 45 percent of code samples failed security tests and introduced OWASP Top 10 vulnerabilities, and that the models failed to defend against cross-site scripting in 86 percent of relevant samples. It also reports that the models got better at writing code that works but no better at writing secure code, and that security results stayed flat regardless of model size. Veracode is a security testing vendor, and this summary comes from its blog post, not the full report.
Dependencies are a second risk. A 2024 study of package hallucinations generated 576,000 code samples with 16 models. The average share of hallucinated package names was at least 5.2 percent for commercial models and 21.7 percent for open-source models. The authors call this a new form of package confusion attack, so check that every dependency exists and is the one you meant before you install it.
Code nobody can maintain
Speed now can cost time later. In Stack Overflow's 2025 survey, 66 percent of developers named "AI solutions that are almost right, but not quite" as a frustration, and 45 percent said debugging AI-generated code takes more time. More respondents distrusted the accuracy of AI tools (46 percent) than trusted it (33 percent). When nobody on the team has read the code, every later change starts with someone learning it.
Scanners help, but they do not sign off
Some builders now include security scanners, and they are worth running. Lovable's docs describe a Quick scan and a Deep scan, and say these tools "do not replace a thorough security review". They also say the app owner is responsible for meeting the security needs of the use case, and suggest a professional security review for apps that handle sensitive data. The researchers behind the CVE argued in May 2025 that the scanner checked only whether a row-level security policy existed, not whether it was correct. That claim is about the scanner as it was then.
What to fix before real users arrive
Fix access first, because that is where one user's data reaches another. This is the short version. Our pre-launch checklist for vibe-coded apps has ten checks with the exact test to run for each.
- Run a two-account test. With the first account's session, request the second account's records by ID on every endpoint, then call each endpoint with no session at all. Nothing private should come back.
- Cover every table. Turn on row-level security for every table the browser can reach, revoke grants the anonymous role does not need, and check views and any policy written as
using (true). - Move the keys. Search the shipped bundle and the repository history for keys, move them to server-side functions and rotate anything that leaked.
- Set payments and roles on the server. Payment status and user roles should change only in server-side code, such as the payment provider's webhook, never in a request the browser can edit.
- Check dependencies. Confirm each package exists, is the one you meant and has a pinned version.
- Prove recovery. Restore a backup into a clean project once, and make sure a failed login and a server error each reach a person.
- Find a reader. Ask someone who did not write the prompts to explain, from the code, who can read and write each table. If nobody can, that is your answer.
Checks 1, 2 and 4 are access control problems, listed as A01 Broken Access Control in the OWASP Top 10:2025, and check 5 falls under A03 Software Supply Chain Failures.
Fix it or rebuild it
Whether to harden or rebuild depends on where the failures sit. Harden the app when the failures are in configuration, keys, logging or a handful of endpoints. Rebuild the data and permission layer, and keep the screens, when records have no clear owner or every route checks permissions its own way. The same goes for moving a Lovable or Bolt prototype to a back end you control: the data and permission layer is the part that needs the most care. This is our rule of thumb from reading prototype code, not a measured threshold.
Getting a second pair of eyes
If you want someone else to read the code, our MVP codebase audit covers architecture, code quality, security and privacy (auth, sessions, role-based access, secrets management and the OWASP Top 10) and reliability. If the answer is a rebuild, the way we work is a free strategy call, then a one-week scope sprint that produces one fixed price, then a 2 to 3 week build.
Limits of this page
We read every source on November 11, 2026: Karpathy's post on X, the NVD record and its data feed, the researchers' write-up, Replit's home page, Lovable's security documentation, Supabase's row-level security documentation, Veracode's blog post (not the full report), the arXiv abstract, Stack Overflow's 2025 survey results and the OWASP Top 10:2025 list. We did not test any AI builder, scan any app or audit any project for this page, so nothing here says a particular tool produces unsafe apps. CVE-2025-48757 is disputed, and the write-up's author works for a competing builder. The table and the fix list are our judgment, not security or legal advice.
Sources
- Karpathy's post naming vibe coding, X, February 2, 2025
- CVE-2025-48757, National Vulnerability Database, published May 30, 2025
- Statement on CVE-2025-48757, Matt Palmer, May 29, 2025
- Security overview, Lovable documentation
- Row Level Security, Supabase documentation, updated November 3, 2026
- 2025 GenAI Code Security Report blog post, Veracode, July 30, 2025
- We Have a Package for You!, a study of package hallucinations by code-generating LLMs, Spracklen and others, arXiv, June 2024
- AI section of the 2025 Developer Survey, Stack Overflow
- OWASP Top 10:2025, OWASP
