Back to Explore
NewsVibeFix editorialUpdated Nov 20, 2026

159,000 Lovable Apps Checked: Passport Numbers and 200,000 Private Messages, Visible Without Login

Crissy Field examined 159,041 Lovable apps: 24,664 websites returned database records on first load, 7,403 held email addresses, and passport numbers plus 200,000 private messages were readable without logging in. The flaw is not one app — it is a whole category leaking the same way.

A person types on a laptop showing a DATA BREACH security warning with a padlock icon

No login. No sign-up. No clicking anything. Just open the page, and the data walks out to meet you. German security firm Crissy Field examined 159,041 apps built with Lovable and found a large share of them behaving exactly like that: database records sent straight to the browser as the homepage loaded, and more available with one extra question through the same interface. The most striking cases include over 1,000 records carrying passport numbers, a German forum exposing more than 80,000 personal profiles and over 200,000 chat messages, and 280,000 emails sent to one provider's customers. The timeline matters, so let's get it straight first: Handelsblatt journalist Lina Knees first reported the investigation on November 2, 2026 (in German); Crissy Field then published an updated English report with newer figures, which reached the Hacker News front page on November 14. The report page itself carries no publication date, so November 14 is the day it drew wide attention — not the day the report was published.

The conclusion up front: this is not one app with a typo in its code. It is a whole category of apps leaking data in the same way. Lovable's promise is "no technical knowledge needed — describe it in words and get a website." What this report punctures is the other half of that sentence: building a website no longer takes technical skill, but protecting data still does, and the responsibility sits with the user. For anyone vibe coding with Lovable, Cursor, v0 or similar tools, this is not distant industry news. It is a mirror: the app you shipped last week may be exposed in exactly this way, and you may have no idea.

How the research was done

Crissy Field is a Hamburg-based security consultancy that does penetration testing and code review. Its method was not exotic: it pulled Lovable projects published between December 2025 and September 2026 from Common Crawl, the public web archive — 159,041 of them — and checked each one for whether it talked to a database as it loaded. The researchers looked at only one database service: Supabase. The report is candid about why: its integration is relatively easy to analyze, and the researchers believe it is the most widely used database service among Lovable projects. Lovable's own database features, introduced in late 2025, are still less widely used and were not included in this analysis — but the report warns explicitly that similar misconfigurations may occur there too.

The research ran in two steps. First, the team watched only what a website sent on its first load — entirely passive, no login, no clicks, no interaction. Second, they used the very same API key the website itself hands out in the page and sent further read-only requests, to see whether more content could be retrieved. That design is the point: it simulates not a sophisticated hacker, but what any ordinary visitor can do — open the developer tools, copy the key that is already sitting in the page, and ask the database again.

The numbers narrow down, and the problem gets concrete

Of the 159,041 projects, 34,545 used Supabase — about 21.7%. Of those, 24,664 websites sent database records from Supabase as soon as they loaded, without requiring a login or any other interaction. Before treating that number as a "leak count," the report itself applies the brakes: data arriving on first load alone is not necessarily a security issue, since it may be intended for public display — a product list, blog posts, event listings. A website sending public content to a visitor is exactly what it is supposed to do.

The real dividing line is step two. Using the API key already present in the page, the researchers sent additional read-only requests and found records containing email addresses in 10,560 tables across 7,403 of the 24,664 websites — 30%. An email address here is a probe: it is the most easily recognizable marker of personal data, and where emails appear, the table almost certainly holds information about people rather than a public product catalogue. In plain terms, at least 7,403 websites had their database doors open to everyone; anyone who opened the page could read the records out.

Funnel chart: of 159,041 Lovable projects, 34,545 used Supabase; 24,664 returned database records on first load; 7,403 contained records with email addresses

What exactly was exposed

The individual cases say more than the totals. The report names no one, but the details are enough to sit up straight. A German lifestyle forum exposed more than 80,000 profiles containing full names, email addresses, dates of birth and gender information, plus more than 200,000 chat messages, all accessible without logging in. Several event organizers and restaurants exposed tens of thousands of ticket and food sales records, including email addresses, phone numbers, postal addresses, payment information and ticket codes — and leaked ticket codes mean someone else can walk in with your ticket.

A property management company exposed AI-generated summaries of internal emails containing sensitive information about tenants, property owners and business partners. The most serious case came from a service provider: more than 1,000 records containing identity document details sat behind a public interface, including passport numbers, expiration dates and nationalities. The report puts it drily: these highly sensitive identity details could be used for identity theft. Beyond that, there were tens of thousands of contracts, invoices, personal photo albums and internal WhatsApp communications, plus large numbers of newsletter subscriptions, waiting lists and contact and support inquiries. In one case, 280,000 emails sent to a provider's customers were accessible. The report neither exaggerates nor softens it: data exposed like this could be used for phishing, ticket fraud or identity theft. Whether any of it has actually been misused, the researchers state plainly, they could not establish.

How not to read the numbers: the report's own limitations

The most trustworthy thing about this report is the full section it spends on how its own numbers should not be read. The 30% figure, the researchers state explicitly, should not be read as a confirmed data leak rate. There are three reasons. First, an email address may be a public contact address or test data — containing an email is not the same as containing a real user's private information. Second, the method misses personal records that contain no email address, so tables holding only names and phone numbers would not be counted. Third, the sample comes from Common Crawl and is not representative of all Lovable applications, and the researchers downloaded only parts of the datasets rather than the full data, so the affected datasets may contain more than the report shows — not less.

In other words, 30% is neither a site-wide leak rate nor a clean bill of health. It holds only within the 24,664 websites examined more closely, and only under the rough probe of "contains an email address." The true scale of exposure could be larger or smaller than that number, but the report's conclusion does not depend on a precise figure: whether the exact value is higher or lower, the fact that "many Lovable applications leave even highly sensitive data unprotected" is already established by a long list of concrete cases. The right way to read this report is to focus on the cases and the mechanism — not to quote 30% around as a site-wide health check.

Why AI-generated apps are especially prone to leaking

Misconfigured database permissions happen in traditional development every day, and the report says so itself: access control mistakes can happen in any application, whether written by people or built with AI. But platforms like Lovable stack three things together that make the same mistake more common, harder to spot, and more concentrated in its consequences.

First, the default configuration leaks. The AI builds the app from your description, connects the database, the frontend can read the data, the feature works, and you ship it. Nobody closes the permissions layer by default, and no step in the flow forces you to confirm "who can read this table." In traditional development, setting database permissions is a step in the launch process that is hard to skip. In vibe coding, that step often does not exist at all, because the whole point of the workflow is to make you worry about less.

Second, nobody checks. The person building the app does not understand database permissions, and the AI will not volunteer that "this table is public right now." The app runs, the pages render, and the user assumes it is safe. One line in the report matters here: users without a technical background face a harder task spotting and preventing these flaws. That is not calling users careless — it points to a structural problem. The platform lowered the barrier to building to zero and left the barrier to building securely exactly where it was.

Third, the patterns repeat, so attackers can sweep in bulk. Database integrations in Lovable apps follow recurring patterns, so an approach that works against one app is likely to work against others with little extra effort. Traditional websites are each written differently, with different weaknesses; AI-generated apps come off the same assembly line, with identical weak spots. That is also what the sample size of 159,041 really proves: not that one app was unlucky, but that a whole category of apps leaks systematically.

And the most uncomfortable part is the gap between promise and responsibility. Lovable's marketing repeatedly emphasizes that no technical knowledge is needed to build projects, which easily leaves the impression that the platform takes care of security too. Yet its Terms & Conditions place responsibility for applications and data handling on the customers themselves. The report's own words: this leaves "a troubling gap between the promise of an easy-to-use, all-in-one service and the responsibilities customers are expected to take on." In plain language: the platform takes responsibility for letting you build it, not for making it safe. Handelsblatt contacted Lovable with the findings; the report does not say how Lovable responded, and no public response from Lovable had been seen at the time of writing.

A pre-launch checklist for vibe coders

If you are building an app with Lovable or a similar tool that handles user data, run through this list before every launch. It is short, it does not require you to become a security expert, and it would have caught most of what this report found.

First, log out and look at your own website again. Open an incognito window, stay logged out, open your app, and check whether things you should not see are visible. Focus on user lists, order records, chat content and profile pages. If you can browse other people's data without logging in, the problem is already confirmed — no one needs to tell you.

Second, check permissions table by table in the database. In Supabase, for example, verify that Row Level Security is actually turned on for every table, and test by hand what the anonymous key can read and write — don't go by feel. When the AI created the tables for you, the permission settings were very likely left open by default, and that default state is exactly the state of the 7,403 websites in this report.

Third, decide who can see what before you launch. Keep public content and user data apart: product lists and event pages can be public; user profiles, orders, chat logs and private messages must be isolated per account. Don't manage every table with one key and one permission set just to save effort. And one hard rule: before permissions have been checked, don't put passport numbers, ID numbers or payment details into the app at all.

Fourth, re-check after major changes. The report's advice is practical: review access permissions, interfaces and separation between user accounts before an application handles real customer or business data, and repeat those checks after significant changes. One AI-driven change of requirements can quietly create a new table with default-open permissions, and what you verified last time may no longer hold.

Fifth, if you don't know, get someone who does to take a look. The report's own line: automated checks are no substitute for a professional security assessment. A platform's built-in scan, or the AI telling you it has "hardened" things, counts as a reference at most — not an acceptance test. Once your app starts taking real users and real orders, it is worth paying a modest sum for a permissions review. That money is far cheaper than apologies and compensation later.

What the platforms should fix

Putting all the blame on users is not fair, and the report does not aim its criticism only at them. There are at least three things platforms can do, none of which requires a technical breakthrough. First, make the default secure: new database tables private by default, sensitive tables requiring login by default, and AI-generated permission policies following least privilege rather than wide open. Users should have to actively choose to publish something, not discover afterwards that they already had.

Second, put the check into the publishing flow. Before the user clicks "publish," run a permissions check and spell out in plain language which tables an anonymous visitor can read and which tables contain email or ID-like fields — then let the user confirm with eyes open. Today's flow ships as soon as the feature works, with security checks left entirely to the user's initiative. That is the soil in which 24,664 websites returned data on first load.

Third, state the responsibility clearly. When marketing says "no technical knowledge needed," it should say in the same breath what the platform handles and what the user has to handle — not bury the key point in the small print of the terms of service. Users are not unwilling to take responsibility; most simply don't know this responsibility is theirs. That information gap is itself the biggest security hole in this report.

What to watch next

Three things are worth watching. First, whether Lovable responds — and whether its response touches the default settings. Handelsblatt has already put the findings to Lovable. A "we are investigating" statement is easy to write; changing defaults and adding pre-publish checks would alter the product flow, and that is the real test of whether the platform is taking it seriously.

Second, how Lovable's own database features hold up. The report only examined Supabase, because it is the most widely used and the easiest to analyze; Lovable's own database features, introduced in late 2025, are still used by few and were not examined. But the same generation logic and the same defaults make it likely the problem exists there in the same form. When someone checks that side, and what they find, is the next result worth waiting for.

Third, whether other platforms get the same treatment. Lovable is not the only "describe it in words, get an app" platform, and any tool that turns database integration into a default template may carry the same weakness. The method in this report is public: Common Crawl plus public interfaces, no special access required. The next platform to be examined this way may well be the one you are using right now.

What is really worth remembering here is not the number 159,000, but the pattern it reveals: AI has cut the barrier to building an app to zero, without cutting the barrier to building a secure app along with it. The only part that disappeared is writing the code. Database permissions, access control and data isolation — the unglamorous work — did not go away at all. It simply moved off the programmer's to-do list and quietly onto the to-do list of users who don't know the first thing about it. Whether your app can launch can now be decided with a single sentence; whether your users' data is safe still depends on a few minutes of checking before you ship. A few minutes of checking permissions is cheap; the money, time and trust a leak costs afterwards is not something a few minutes can buy back. Don't let your app become the anonymous case study in the next report.

Views 9Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A developer with glasses types on a laptop late at night at a wooden desk, a large monitor behind showing colourful code, blurred city lights outside the window
News
Run the Middle Layers Three Times: Inkling 2 Mini Puts a 26B Coding Model on Your Laptop

On 20 November 2026, Thinking Machines Lab released Inkling 2 Mini: a 26.4B-parameter looped transformer that runs its middle 12 layers three times per token, fits in about 20GB, and runs offline on a laptop. It leads open-weights models of its size on DeepSWE and SWE-bench Verified — with the caveats printed in the same announcement.

Product LaunchModel UpdatesAI Coding
A developer with glasses types on a laptop at a wooden desk, an external monitor to the left showing C code on a dark background, a small plant on the desk
News
Microsoft's Coding Model Stops Playing Second Fiddle: MAI Code 2 Scores 96% of Opus 5.5 for 5% of the Cost — Read the Three Caveats First

On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.

Developer WorkflowProduct LaunchModel Updates