AI coding assistants are genuinely great. They are also confidently wrong in ways that are hard to spot, because the output looks like something a competent developer wrote. Security is where this bites hardest.
Is vibe-coded software secure?
Vibe-coded software is not secure by default. In Veracode's 2026 GenAI Code Security Report, AI models wrote code that compiled almost every time but passed security tests only 56% of the time, barely up from 55% a year earlier. The best model tested still failed roughly one security task in three.
The detail inside that report, published in July 2026, is worse than the headline. Results swing by language (Python passed 63% of the time, Java 30%), and models built specifically for coding were no safer than general-purpose ones. Models keep getting better at writing code that works. They are not getting better at writing code that is safe, and they now write a large share of what gets committed.
The consequences are showing up in public vulnerability records. Georgia Tech's Vibe Security Radar project had traced 74 CVEs to AI coding tools by March 2026, with the monthly count climbing from 6 in January to 15 in February and 35 in March. The researchers estimate the true figure is five to ten times higher, because most AI-written code carries no marker saying so.
Plausible but wrong
A model will happily generate auth middleware that decodes a token without verifying its signature, or a database call that concatenates user input into a query. It compiles. It runs. It passes your happy-path test. And it is exploitable.
The clearest real-world case is CVE-2025-48757. A researcher scanned 1,645 apps built with the AI app builder Lovable and found 170 of them exposing their databases to unauthenticated requests, because the generated projects shipped with missing or insufficient Row Level Security policies. From the builder's point of view nothing was broken. The apps worked. That is what this failure looks like: a working product with the door open.
The patterns that keep showing up
jwt.decode()used wherejwt.verify()was needed, trusting a token nobody validated. See JWT authentication mistakes that ship to production.- Secrets with a fallback default like
process.env.SECRET || "secret", or pasted straight into source. See how apps leak API keys. - Firebase or Supabase rules generated wide open because the demo needed it. See Firebase rules are open by default and whether the Supabase anon key is safe to expose.
bcryptrounds set absurdly low, or hashing skipped entirely.- User input dropped straight into SQL, shell commands, or HTML.
- Auth left as a comment: a route with
// TODO: add authabove it and nothing below.
The model did not make a mistake a beginner makes. It made a mistake that looks like expertise, which is worse.
Most of these map directly onto the OWASP Top 10. They are not new bugs. What is new is the rate at which they get written and the confidence with which they get shipped.
Why does code review miss AI-generated vulnerabilities?
Code review misses AI-generated vulnerabilities because the code looks finished. When you did not write it, you read it more charitably and skim. The insecure line is often one token away from the secure one, decode where verify belonged, and nothing fails when you run it, so there is no error to draw your eye.
There is a second problem: you never made the decision. When you write auth by hand, you choose how tokens get checked, and you remember choosing. When a model writes it, the choice was made for you, silently, and you only learn which way it went if you go looking.
Scan for the specific patterns
You cannot eyeball your way to safety across a whole AI-assisted codebase. What works is scanning for the specific known-bad patterns that AI tools produce, by name, and flagging each one with the fix. That turns a vague worry into a concrete checklist.
If you are close to launch, start with the pre-launch security checklist, six checks that take under ten minutes each, or the shorter 60-second pre-launch audit. Then check what you have already committed, because a secret deleted from the current code is still sitting in your git history.
IOnclad automates this pass. Its Vibe Code Pack looks for the AI-specific patterns above alongside hardcoded secrets, OWASP Top 10 issues, and git history. It runs offline, so your source never leaves your machine, and it ends in one verdict: Ship It or Not Ready.
This post is part of our app security guides for indie developers.