# One Lovable forum exposed over 80,000 profiles without a login

**Summary:** Crissy Field's Lovable study documents exposed private records, showing why a working app still needs a review of database permissions before handling sensitive data.

- Canonical: https://markhuang.ai/news/lovable-forum-exposed-profiles
- Language: en
- Author: [Mark Huang](https://markhuang.ai/about)
- Published: 2026-11-14
- Section: News
- Tags: Lovable, Supabase, AI App Development, Access Control, Data Privacy
- Source: [Crissy Field](https://www.crissyfield.de/research/lovable-en/)
- License: https://creativecommons.org/licenses/by-nc/4.0/

---

![A polished cabinet with coral and violet front panels has an open side drawer full of blank profile cards](https://cdn.markhuang.ai/news/lovable-forum-exposed-profiles/hero.webp)

*Conceptual illustration: an app's polished surface says little about who can reach the records behind it. This is not an image of an affected system.*

A German lifestyle forum built with Lovable exposed more than 80,000 profiles, including names, email addresses, birth dates and gender information. More than 200,000 chat messages were accessible too. Researchers at Crissy Field say they could retrieve this information without logging in, according to [their investigation of Lovable applications](https://www.crissyfield.de/research/lovable-en/).

Lovable turns plain-language descriptions into web applications, including apps with accounts and databases. For a builder, a successful demonstration therefore leaves another question unanswered: who can read the data behind it?

For someone turning an idea into a customer service or an internal tool, the useful next step is a different kind of review. Alongside checking what the app lets the intended user do, check what it refuses to let everyone else do. That becomes especially important before putting real people's information into it.

## The exposed records included AI summaries

The forum was one of several serious cases Crissy Field describes. A property management company exposed AI-generated summaries of internal emails containing sensitive information about tenants, property owners and business partners. Another service provider exposed more than 1,000 records containing identity document details, including passport numbers and expiration dates.

Turning an email into a shorter AI-written digest does not remove the need to protect the information it contains. A review of a tool that summarizes business correspondence needs to cover the saved summaries as well as the original messages.

The researchers retrieved only portions of the datasets and did not establish whether anyone had misused the exposed information. Accessible records demonstrate a privacy failure; they do not, by themselves, establish subsequent fraud or identity theft.

## What the study's 30% actually measures

Crissy Field identified 159,041 Lovable projects using the Common Crawl web archive and analyzed them between December 2025 and September 2026. The investigation focused on Supabase, a database service used by 34,545 of those projects. It did not examine Lovable's own database offering.

Of the Supabase projects, 24,664 websites sent database records when their pages first loaded, without a login or other interaction. That can be legitimate: a public website needs to deliver public information. The researchers then made additional read-only requests using the same API key and found records containing email addresses in 7,403 of those websites, across 10,560 database tables.

That is the source of the 30% figure. It describes email-containing records in a selected group of 24,664 websites, not confirmed leaks across all Lovable apps. Public contact addresses and test data can match that signal; private records without email addresses can escape it. Common Crawl also does not provide a representative sample of every Lovable application.

I find the documented sensitive records more useful than a headline percentage here. They establish that serious exposure occurred without pretending the study measured how likely any particular Lovable app is to leak.

## A login answers only part of the question

Signing in establishes who a user is. Access rules determine what that user, or someone who has not signed in, may read or change. A page that hides a record is not enough if the service supplying that record still returns it to an unauthorized request.

[Supabase's row-level security documentation](https://supabase.com/docs/guides/database/postgres/row-level-security) describes rules enforced inside the database that limit which records a request can access. One example allows a user to read their own to-do items. Table permissions also matter: the service must control both the operations someone may perform and the records those operations apply to.

This simplified relationship explains why reviewing the visible page alone cannot settle the question. It illustrates the access boundary, rather than reconstructing a particular app from the investigation:

```mermaid
flowchart TD
  A["Visitor uses the app page"] --> B["Request for database records"]
  C["Visitor requests data outside the page"] --> B
  B --> D["Database permissions and record access rules"]
  D --> E["Return only records this visitor may read"]
```

Those rules are a documented safeguard, not a diagnosis of every app in Crissy Field's report. The investigation does not supply enough detail to attribute every exposure to one specific configuration mistake.

## Give the review something specific to prove

Lovable already provides security checks. Its [current documentation](https://docs.lovable.dev/features/security) describes a Quick scan that checks database access rules, among other things, and a Deep scan that reviews application logic and permissions. The same documentation says the tools cannot guarantee complete security and do not replace a thorough review. Those current capabilities do not establish which checks ran on the apps in the study.

My recommendation is to write down who should be allowed to see or change each kind of record, then ask a qualified reviewer to check those rules. Using synthetic records in an environment you control, check that a signed-out visitor cannot retrieve private information. Then check that one ordinary account cannot read another account's private records. Confirm that the intended user can still do the work the app promises. These are review cases, not a complete security assessment.

Crissy Field recommends reviewing permissions, interfaces and separation between accounts before an app handles real customer or business data, then repeating those checks after significant changes.

This is a concrete version of the learning problem I wrote about in [AI Makes Building Easier. I Still Have to Learn.](/blog/ai-makes-building-easier) AI can help someone produce a useful tool before they understand every decision inside it. Before that tool holds private information, its owner needs evidence that someone without permission cannot read it.
