Applications built with the AI coding platform Lovable have been serving passport numbers, private chat messages, invoices and payment records to anyone who loaded the right page. A research report by Crissy Field, an application-security consultancy in Hamburg, identified 159,041 Lovable projects in public web-archive data from December 2025 to September 2026. Of these, 34,545 used the backend database service Supabase, and 24,664 websites sent database records as soon as they loaded, without requiring a login or any interaction. In records pulled from those 24,664 sites, email addresses turned up in 10,560 tables across 7,403 of them, about 30 percent. "We found many clear examples of sensitive data left unprotected," the report states. [1]
Nothing in the report describes a break-in. The data was reachable through the applications' own configuration, and the exposure extends to the platform's builders generally, beyond the sites on the report's list. [1] The practical response is a configuration check, and the tools for it cost nothing to run. [2]
For a builder or a company operating an app on Lovable, the response comes down to five checks, and the first three are where to start. Look at the security findings that publish-time scanning already produces, run the free deep scan before pointing an app at real customer data, and keep government identifiers such as passport numbers out of standard-plan apps unless your agreement with Lovable explicitly allows them. [2][3][4]
24,664 sites answered page loads with database records
The scale is documented, and every headline number in it comes with a stated limit. Crissy Field worked from Common Crawl, the open web archive, covering December 2025 to September 2026. It identified 159,041 Lovable projects, found that 34,545 of them used Supabase, and counted 24,664 websites that sent database records on first load with no login or other interaction. [1]
The report is explicit that the raw count needs its conditions. "That alone is not a security issue: the data may be intended for public display," it says of the first-load records. [1] The further content came from read-only requests made with the same API key, and the team reports that "The 30% figure should therefore not be read as a confirmed data leak rate. It applies only to the 24,664 websites we examined more closely," with a sample that is not representative of all Lovable applications. [1]
The figures cover only the records the researchers actually downloaded, so the affected datasets may hold more, and what they downloaded was concrete. A German lifestyle forum exposed more than 80,000 profiles with full names, email addresses, dates of birth and gender information, alongside more than 200,000 chat messages. Several event organizers and restaurants exposed tens of thousands of ticket and food-sale records, including payment information and ticket codes. One service provider exposed more than 1,000 records containing identity document details, among them passport numbers, expiration dates and nationalities. The list also includes 280,000 emails sent to one provider's customers, and tens of thousands of contracts, invoices, personal photo albums and internal WhatsApp communications. On consequences, the researchers stay inside their evidence. "We could not establish whether any of it had been misused," they write. [1]
The investigation was carried out with Handelsblatt, which reported the findings on November 2, 2026, and the report states that its own figures are more recent than that article's. [1] By November 15 the story sat on the Hacker News front page with 638 points and 271 comments. [5]
Row-level security left open made the exposure repeatable
Row-level security is the database rule that decides which rows a client may read. When an application ships with it left open, the records that fill a page also answer further queries, which is what the researchers' read-only requests did with the same API key the apps already use. Lovable's head of security attributed the root cause to insufficient row-level security and tied it to the company's own tooling, writing that it is "exactly what our scanning is built to catch." [2]
The repeatable part matters more than any single site. "With Lovable, database integrations follow recurring patterns, making it easier for attackers to look for similar misconfigurations across many applications," the report says, adding that users without a technical background face a harder task spotting and preventing these flaws. [1] The research focused on Supabase, which the team believes is the most widely used database service among Lovable projects; Lovable's own database features, introduced in late 2025, were not examined, and the report says similar misconfigurations may occur there too. [1] An app can be built today, outside the examined sample, with the same wiring and the same exposure, and the publish-time scan and the deep scan are the checks that catch it. [2][3]
Lovable's answer puts the final call with the builder
Lovable's written response, posted in the discussion by its head of security, is that every app is scanned automatically before publish and a deeper code scan, static application security testing (SAST), is offered, "both free for all users." Flagged findings can be sent to the Fix option for resolution, auto-fix can resolve security findings in the background, and publishing can be blocked entirely when critical findings exist. [2]
The company's June 1, 2026 announcement describes the mechanics. The publish-time scan runs "in about 10 to 15 seconds" and "It checks for common issues, including database misconfigurations, missing RLS policies, and authorization gaps." The basic scan "doesn't analyze your application code for logic flaws"; the deep scan is "a full AI-powered review of your codebase" taking "about 2-4 minutes," it is "not required to publish," and "Auto-fix is opt-in." [3] A November 10, 2026 post signed by the same head of security adds that findings are shown before an app goes live and that workspace admins can block publishing on critical findings. The same post states that "your app is yours, and so is the final call on what goes live." [6]
The terms draw the same line in contract language. Under the version dated August 28, 2026, effective August 15, 2026, "You are solely responsible for reviewing, validating, and using any AI Output" and "You are responsible for your Customer Data and for the applications and projects you build." A No Sensitive Data clause excludes government identifiers such as passport numbers, and payment card data, from the standard services unless a plan, a data processing agreement or an enterprise agreement explicitly permits them. [4] The same company's marketing tells builders they can "Build a proof-of-concept without writing a single line of code. If you can describe it, you can build it." [7] Crissy Field frames the distance between those two positions as "a troubling gap between the promise of an easy-to-use, all-in-one service and the responsibilities customers are expected to take on." [1]
Two open points sit inside this picture. The automatic publish-time scan became the default on June 1, 2026, while the period the researchers examined starts in December 2025; the report does not say when the affected sites were published or whether Lovable's scanner had flagged them, so no conclusion can be drawn in either direction. And the discussion itself split, with one commenter judging that "Lovable is pretty much untouchable for serious businesses going forward until they clean up this mess" [8] and others arguing that opt-in auto-fix is no substitute for a framework around how sensitive data is collected and used. [5] Those are positions in the thread, not established facts. [5]
Five checks before real data goes into a Lovable app
The first two checks read output Lovable's scans already produce; the last three are the builder's own work — a data match, a set of tooling choices, and a permissions review.
-
Read the findings your app already produces at publish time. Scanning runs automatically and flagged issues include missing RLS policies and authorization gaps. [2][3] The basic scan takes about 10 to 15 seconds and does not analyze application code for logic flaws. [3]
-
Run the deep scan before handling real customer data. It reviews the full codebase in about 2-4 minutes, is free to run anytime, and is not required to publish. [3]
-
Match your data model to the terms. Standard plans exclude government identifiers and payment card data unless your agreement allows them, so an app already holding passport records has a terms problem and a security problem at once, and removing the data comes before anything else. [4]
-
Decide deliberately on auto-fix and publish blocking. Auto-fix resolves straightforward, low-risk findings only when switched on, and admins can set publishing to block on critical findings for workspaces where that matters. [2][3][6]
-
Review access permissions, interfaces and the separation between user accounts before real customer or business data goes in, and repeat the review after significant changes. The report states that "Automated checks are no substitute for a professional security assessment." It also advises bringing in qualified support when the expertise is not available in-house. [1]
The documented exposure in this event sits at the settings layer, where the platform's free tooling meets the builder's decisions. Lovable's position, in its discussion reply and its own documentation, is that basic scanning is automatic, the deep scan is free, and the final call belongs to the builder. [2][6] The report cannot say whether the affected apps shipped before or after scanning became a default in June 2026, and it could not establish that the data was misused. [1] The five checks split by cost: the two automated ones take seconds to a few minutes, and the terms match, the tooling choices and the permissions review are the work that remains. The one cost that cannot be undone is data that should never have entered a standard-plan app.