Anyone can now describe an app to an AI tool and have something working by the afternoon. That is genuinely useful, and it is also why some businesses are being offered very cheap apps that nobody has properly checked. The risk is not that AI wrote the code. The risk is that nobody with security experience reviewed it before real customer data went in.

This article explains the most common weaknesses in plain language, what recent research actually found, and the questions to ask before you trust an AI-built app, or the cheapest person who offers to build one.

In short

  • "Vibe coding" means building software by prompting AI tools, often with little or no manual code review.
  • Researchers keep finding the same basic gaps: databases left open, secrets visible in the page, and missing login or permission checks.
  • AI-assisted development is not the problem. Shipping without review, testing and an owner after launch is.
  • Before you buy, ask who reviews the code, how permissions are tested, where your data lives and who is responsible after launch.

What is vibe coding?

Vibe coding is building an app by describing what you want to an AI tool and accepting what it produces, rather than writing and checking each line yourself. The term is now used for everything from a weekend prototype to a tool a small company relies on every day.

Used by a skilled developer who reads and tests the output, it can save real time. Used by someone who cannot judge the result, it can produce an app that works in a demo but has never been tested against deliberate misuse.

Are AI-built apps safe?

Not automatically, and the evidence suggests many are not, at least when they ship without review. No single study settles it, so here are the findings I could confirm from their sources, with their limits.

  • Veracode tested code from more than 100 large language models and reported that 45% of the samples introduced OWASP Top 10 vulnerabilities. It also reported that security did not improve with larger models.
  • Tenzai had five AI coding tools each build three apps from the same prompts (15 in total) in December 2025. As InfoWorld reported, it found 69 vulnerabilities, about six of them critical. The tools avoided flaws with a standard fix, such as SQL injection and cross-site scripting, but struggled with context-dependent problems like authorisation logic.
  • RedAccess, an Israeli security firm, reported finding about 380,000 publicly reachable apps and assets built with tools such as Lovable, Base44, Replit and Netlify, around 5,000 of which exposed sensitive corporate information, according to Security Boulevard. The tool makers disputed parts of this. Some said a publicly reachable app is not automatically a breach, and Base44 and Lovable said the URLs needed to verify the findings were not supplied. Treat it as a warning, not a verdict.
  • Moltbook, a social network for AI agents, left its production database open. Security firm Wiz published its findings on 2 February 2026: a database key sat in the page's JavaScript and Row Level Security (the rules limiting who can read or write each record) was missing. Wiz reported about 1.5 million API tokens and around 35,000 email addresses exposed, and Moltbook fixed it within about a day.
  • Georgia Tech's Vibe Security Radar, as summarised in a Cloud Security Alliance research note from April 2026, had confirmed 74 AI-linked CVEs (publicly catalogued vulnerabilities) through March 2026, rising from 6 new ones in January to 35 in March. The researchers believe the true number is higher, but that is an estimate.

A widely shared claim that over 90% of vibe-coded apps are vulnerable could not be traced to a primary source, so I have not used it.

What goes wrong most often?

The same handful of weaknesses appear again and again. None needs a security degree to understand.

Risk What it looks like What to ask
No login or permission checks A staff member, or a stranger who guesses a link, can open other people's records or admin pages "How do you test that one user cannot see another user's data?"
Database left open to the public The app talks directly to the database, and nobody set rules for who can read or write (the Moltbook case) "Who can read and write the database, and where are those rules written?"
Secrets exposed in the page Passwords or API keys sit in code that every visitor's browser downloads "Where are keys and passwords stored, and who can see them?"
No input validation Forms accept anything, so attackers can send harmful data or make the server fetch things it should not "What happens if someone submits unexpected data?"
No security headers or rate limits The app has no browser protections and no limit on repeated guessing "What protects against repeated login attempts and common web attacks?"
No backups One mistake, bad update or attack and the data is gone "Where are backups stored, and when did you last test a restore?"

AI tools are good at making features work and weaker at deciding who is allowed to do what.

What should I check before launching an AI-built app?

Use this list before any real customer or staff data goes in. Ask for answers in writing.

  1. Review. A named person with security experience has read the code, not only run it.
  2. Permissions tested. Someone has tried to open another user's data and failed. Ask to see how.
  3. Data location. You know which service and country hold your data, and who has admin access.
  4. Secrets. No passwords or keys are visible in the page or stored in shared documents.
  5. Backups. They run automatically, are stored away from the app, and a restore has been tested.
  6. Updates. Someone is responsible for patching the app and the services it depends on.
  7. Privacy settings. Test accounts and sample data are removed, and the app is not public by accident.
  8. Exit plan. You can export your data and keep ownership of the code, domain and accounts.

Questions to ask your web person

  • Who reviews the code before it goes live, and is that person different from the one who built it?
  • How do you test logins and permissions, and can you show me the result?
  • Where is our data stored, and who else can access it?
  • Who is responsible for fixes, updates and security after launch, and for how long?
  • How often are backups made, and when did you last restore one?
  • If we stop working together, what do we keep: the code, the data, the hosting and the domain?

A good developer will answer these calmly. Vague answers are information.

Does Malaysian law care about this?

If your app holds customers' personal data, it is likely relevant to Malaysia's Personal Data Protection Act 2010. The Personal Data Protection Commissioner's site, pdp.gov.my, describes seven data protection principles and has a section for reporting data breach notifications following amendments to the Act. I am not a lawyer and this is not legal advice. If you handle customer data at scale, check the current requirements with the Commissioner's office or a qualified adviser.

When is a proper business system worth it?

A cheap or AI-generated tool can be a sensible way to test an idea with sample data. It becomes a poor choice when the app holds real customer, member or financial records, handles payments or bookings, or becomes something the business cannot run without.

At that point you are paying for more than code. You are paying for discovery (understanding how your business actually works), design, review and testing, backups, and someone accountable after launch. Those are what turn an app that works in a demo into a system you can rely on.

Frequently asked questions

Should I avoid AI-built apps completely? No. AI tools are useful, including for experienced developers. What matters is that a skilled person reviews and tests the result, and someone is accountable after launch.

Is a cheaper builder always less secure? Not always, but price often reflects how much review, testing and support is included. Ask what the price covers, and do not assume the work includes security checks unless it says so.

What if my app is already live? Start with the checklist, then ask a professional to review it. If it holds customer data, confirm where backups are and who has access, and fix anything exposed first.

Where I can help

I am Rahim, founder of Chillo Creative. I use modern tools, including AI, in my work, but what you are buying is a result you can rely on: a system built around how your business works, reviewed and tested before launch. If you are unsure about an app or a quote you have received, the free Digital Business Audit is a calm place to start, with no pressure. If you need something more than a quick tool, my custom business systems are built after discovery and include 12-Month Post-Launch Support, so someone is accountable after launch.

Sources