Vibe Coding Security Risks: Lovable Apps Exposed 200,000 Chat Messages, the OWASP Top 10 Mistakes Behind It, and What to Check This Week
Researchers found thousands of apps built with Lovable sending database records to anyone who loads the page, including 200,000 chat messages and passport details. Here is what is confirmed, how it maps to the OWASP Top 10, and what to check.
Where Germany · Europe, Western Europe, Worldwide
On this page 8 sections
Researchers have found thousands of apps built with the AI app builder Lovable sending database records to anyone who opens them, including more than 200,000 chat messages and over 1,000 records of passport details. If your business runs an app or internal tool that an AI wrote and no person reviewed, the same mistakes could be exposing your customers' data today.
The failures are not new or exotic. They line up with the OWASP Top 10, the standard list of web app risks, and most of them can be checked this week.
What researchers found in Lovable apps
Crissy Field GmbH, a Hamburg firm that works on protecting applications from attack, ran the study with the German business paper Handelsblatt, which published its story on November 2, 2026. The team sampled 159,041 Lovable projects found in Common Crawl, a public archive of the web, between December 2025 and September 2026.
What the report confirms:
- 34,545 of the projects used Supabase, the hosted database many Lovable apps are built on.
- 24,664 websites sent Supabase database records as the page loaded, with no login.
- Email addresses appeared in 10,560 tables across 7,403 of those websites.
- A German lifestyle forum exposed more than 80,000 profiles and more than 200,000 chat messages.
- A service provider exposed more than 1,000 records of identity document details, including passport numbers, expiration dates and nationalities.
- Other exposed data included ticket and food sales records, contracts, invoices, personal photo albums, internal WhatsApp messages and 280,000 emails one provider had sent its customers.
What is not known:
- Whether anyone misused the data. The researchers write: "We could not establish whether any of it had been misused."
- How common real leaks are. The researchers say their 30% email figure "should therefore not be read as a confirmed data leak rate." Some records that apps send without a login are meant to be public.
- How Lovable responded. The report says Handelsblatt put the findings to Lovable but does not describe a reply.
One detail to get right: the report describes passport numbers and related details, not scanned images of passports.
The earlier warning: CVE-2025-48757
This pattern has been documented before. Researcher Matt Palmer found the issue on March 20, 2025, told Lovable the next day and disclosed it publicly on May 29, 2025. The NVD entry, published May 30, 2025, describes "an insufficient database Row-Level Security policy in Lovable through 2025-04-15" that lets attackers with no login read and modify tables in sites built on the platform. It carries a CVSS 3.1 score of 9.3, critical, assigned by MITRE.
Palmer's write-up explains the cause. The generated apps talk to the database straight from the visitor's browser using a public key, and depend entirely on Supabase Row Level Security (RLS) rules to keep each person to their own rows. When RLS is off or too loose, that public key opens the table. He listed names, email addresses, third-party API keys and payment records among what could be reached.
Lovable disputes the CVE. Its position, recorded in the NVD entry, is that each customer is responsible for protecting its own application's data.
Who is responsible for security in an AI-built app
On that point Lovable is consistent. The Crissy Field report notes that Lovable's terms "place responsibility for applications and data handling on the customers themselves."
That is the business risk in one sentence. The tool wrote the code, but the data, the customers, any breach notification duty and the reputation all belong to the company that published the app. Our view is simple: AI can write code, and a human must review it before it ships.
Vibe coding vulnerabilities mapped to the OWASP Top 10
The OWASP Top 10 2025 is the current edition of the list. Here is where AI-built apps tend to fail, and which failures these reports actually documented.
| OWASP 2025 category | How it shows up in AI-built apps | In the Lovable reports? |
|---|---|---|
| A01 Broken Access Control | Database rules missing or too loose, so one visitor can read everyone's records | Yes, the root cause of CVE-2025-48757 |
| A02 Security Misconfiguration | Protections that must be switched on and written per table never are | Yes, RLS must be enabled and configured for each table |
| A03 Software Supply Chain Failures | Packages the AI added that nobody vetted or updates | Not covered, check anyway |
| A05 Injection | User input passed straight into queries or commands | Not covered |
| A07 Authentication Failures | Login and password reset flows nobody tested | Not covered |
| A09 Security Logging and Alerting Failures | No record of who read what, so nobody can say what leaked | Relevant: misuse could not be established either way |
An AI builder produces something that works when you click through it. Whether a stranger can do things they should not is a separate question, and the app looks just as finished either way.
What to do this week
| Step | How to check |
|---|---|
| List your AI-built apps | Ask every team which tools they published with Lovable or similar builders, and what data each one touches |
| Check Row Level Security | In each Supabase project, confirm RLS is enabled on every table and each policy limits rows to the signed-in user |
| Test as a stranger | Open the app logged out in a private window, open the browser's developer tools and watch the network tab; if records come back, anyone can get them |
| Move secrets off the browser | No API keys in front-end code or readable tables; rotate any key that was ever exposed |
| Turn on access logs | Make sure you could answer "who read this table last month?" |
| Get a human review | Have someone qualified review access controls before real customer data goes in, and again after major changes |
| If you find an exposure | Close it, rotate keys, keep the logs and check your notification duties with counsel |
The researchers say it directly: "Automated checks are no substitute for a professional security assessment."
Why it matters for a company network or website
AI builders let anyone in a company publish a working app connected to real data: a quote calculator, an intake form, a client portal. Many of these never pass through IT, so nobody knows they exist until something goes wrong. Each one is a public door into your data, and its lock is only as good as rules nobody may have written.
Where the app runs matters too. On infrastructure you control, you decide who can reach the database and you keep the logs; our self-hosting infrastructure page explains what that involves.
Vibe coding security questions people are asking
Is vibe coding safe for business apps?
It can be, but not unreviewed. AI-generated code needs the same security review as code a person wrote, especially the access rules around customer data.
Was my data exposed in the Lovable research?
The report does not publish a list of affected apps. If you use an app built on Lovable, ask its owner whether they have checked their database access rules.
Did Lovable fix CVE-2025-48757?
Palmer's disclosure lists a patch release on April 24, 2025 and says Lovable's Publish feature helps make sure RLS is enabled. Enabled is not the same as correct, and Lovable disputes the CVE, so each app still needs its own policies checked.
What is Row Level Security?
A database feature that decides, row by row, which records each user may read or change. Apps that call the database from the browser depend on it completely.
Is AI-generated code less secure than human code?
No single number answers that. These findings show the failure that matters most: code that runs fine but was never checked for who else can use it.
What are the biggest vibe coding dangers?
Broken access control, secrets left in the browser, unvetted packages and missing logs. The first is behind both Lovable reports.
The bigger lesson
A working app and a safe app look identical on screen. The difference only shows when someone asks the database directly, and attackers ask. Put a person between the AI and your customers' data.
Need a second set of eyes on an AI-built app? See our security hardening services or get a free quote.
Get security alerts by email, with the free checklist, at our security alerts page.
Sources: Crissy Field: Lovable research; NVD: CVE-2025-48757; Matt Palmer: CVE-2025-48757; OWASP Top 10.
Have a question about this article?
Ask it here and we will answer you directly by email. Questions are private, never posted.
Security alerts by email
Hear about the next one first.
When attackers start using a flaw in software businesses run on, we email what it is, whether it touches you and what to do that week, with a round-up of our latest security writing. Free, one-click unsubscribe. You get the checklist straight away: 12 things to secure this week.
Want a result like this?
Security alerts, breaches and the technology a business runs on: what happened, who it affects and what to do.
Get a Tech QuoteKeep reading
Moldova Cyberattack: State Actor Suspected After Citizen Data Was Taken From Government Systems, What Is Confirmed and What to Check
Moldova says hackers accessed the personal data of about 450,000 citizens through government digital services....
Park Dental Data Breach Settlement: Up to $2,500 for Patients, a December 28 Claim Deadline, and Why the Case Is Still Not Over
Park Dental patients have until December 28, 2026 to claim up to $2,500 or $60, plus credit monitoring, over a...
Rosstat Cyberattack: Russia's Statistics Agency Is Still Restoring Systems, What Is Confirmed and What Businesses Should Check
Rosstat says it is still restoring systems after cyberattacks from October 29 to November 1, 2026. Hackers cla...
