Software generators leave database doors open to anyone who asks
AI platforms allow beginners to assemble full web applications from plain text, but automated database setups can expose personal documents and conversations when basic access controls are forgotten.

When someone tells an artificial intelligence tool to build an application, the software can assemble web pages, wire up user accounts, and attach an external database within minutes. The convenience hides an unspoken assumption about how modern websites protect private information. Building a functional digital service usually requires two distinct tasks: displaying data to visitors and deciding exactly who is permitted to view each record. When an automated platform handles both, the resulting code can easily confuse public visibility with authorized access, leaving sensitive files open to anyone who queries the underlying storage system directly.
The mechanics of modern web storage explain how that exposure occurs. Most web applications keep their information in separate databases that communicate over standard internet protocols.1 To let a visitor's browser fetch necessary interface elements, like a company logo or a public product catalogue, the website hands the browser a public identification key.1 This key functions like a library pass: it identifies which building the visitor entered, but it is not meant to unlock private offices. Keeping private records safe depends on a second layer of defense known as row-level security, a set of granular rules inside the database that inspects who is asking and decides whether that specific user owns the requested row.1 If an application creator enables the public key but forgets to turn on row-level security, or writes a rule that approves every request unconditionally, the database answers any query sent with the public key.1 Anyone inspecting the code delivered to their browser can take that public key, send a read request directly to the storage server, and retrieve whatever tables sit behind it.1
That disconnect between effortless assembly and technical security has surfaced across thousands of software projects generated through automated coding assistants. Security researchers at the German firm Crissy Field examined archives of web applications built with Lovable, an artificial intelligence platform designed to turn everyday written instructions into working software.23 Across a sample of projects collected between late 2025 and autumn 2026, the researchers discovered that thousands of public applications were exposing personal files, private communications, and identity documents to any browser that requested them.21
How did private records become accessible without a password?
Database records became accessible because applications delivered functional public connection keys without restricting the underlying tables.1 Crissy Field examined 159,041 projects created through Lovable and preserved in the Common Crawl web archive from December 2025 to September 2026.21 The researchers narrowed their focus to projects using Supabase, a widely adopted database provider that connects to Lovable applications.23 Among the archived projects, 34,545 connected to Supabase, and 24,664 of those websites sent database records immediately upon loading in a browser, without asking the visitor to sign in or complete any authentication.23

Serving initial records on first load does not constitute a flaw on its own, because many websites deliberately show public content to anonymous visitors.23 The vulnerability appeared when the researchers took the public connection keys embedded in those pages and issued additional read-only requests for other tables stored in the same databases.21 Because the underlying projects lacked restrictive row-level security policies, the databases answered those queries as well. The researchers found tables containing email addresses across 7,403 separate websites, representing 30 percent of the 24,664 applications that loaded database records without authentication.21
The exposed material went far beyond test text or public email addresses. In one German lifestyle discussion forum, researchers uncovered more than 80,000 user profiles containing full names, email addresses, dates of birth, and gender designations, alongside more than 200,000 private chat messages.21 Several event organizers and dining businesses left tens of thousands of purchase records exposed, including customer phone numbers, postal addresses, payment details, and valid admission ticket codes.21 Another service provider left more than 1,000 identity records readable without authentication, exposing passport numbers, expiration dates, and nationalities.23 Additional exposed databases contained contracts, commercial invoices, personal photo collections, AI-generated summaries of internal property management emails, and 280,000 customer notifications from a single provider.1
Why does natural language coding create this vulnerability?
Natural language coding tools create structural vulnerabilities by encouraging non-programmers to assemble complex software while shifting administrative responsibilities onto the user.2 Anton Osika and Fabian Hedin founded Lovable in Stockholm in 2023, developing the company from an open-source tool called GPT Engineer into an autonomous web generation platform that launched publicly in November 2024.2 Marketing for the service highlighted that individuals required no traditional programming knowledge to deploy working software, but the company's terms of service assigned full legal and administrative responsibility for data protection to the customers who generated each application.23
Because automated generation systems rely on consistent templates to connect front-end interfaces to back-end databases, misconfigurations repeat in predictable patterns.2 When human programmers write software manually, errors tend to be idiosyncratic. In contrast, an automated assistant that generates database connection logic using identical boilerplate can replicate the same permissive defaults across thousands of independent domains.1 An unauthorized party who discovers an open configuration in one generated project can apply the exact same query technique against other applications built on the same architecture.21

The security team at Lovable acknowledged that row-level security misconfigurations were the root cause of the exposures.1 The company explained that its development pipeline includes automatic pre-publication security scans, free static analysis tooling, automated background remediation, and publication blocks for critical configuration errors.1 Lovable also noted that unsecured database access represents an industry-wide integration challenge rather than an internal flaw in its generative neural networks, pointing out that developers encounter similar configuration difficulties across multiple modern coding environments.13
What can the investigation tell us about total exposure?
The investigation provides a detailed sample of open database tables but does not establish a definitive leak rate across all software generated by artificial intelligence.23 Crissy Field based its research on public snapshots gathered from the Common Crawl internet archive, which represents an opportunistic collection of accessible websites rather than a statistically representative sample of all applications created through Lovable.1 The 30 percent calculation applies strictly to the 24,664 Supabase-connected websites that the researchers inspected directly after observing unauthenticated data on first load.21
The methodology used to identify personal records also carried inherent measurement uncertainties. The researchers used the presence of email addresses as an automated indicator of personal information, which can produce false signals if the addresses belong to generic support inboxes or artificial test entries.21 Conversely, tables containing real customer information without an email field were omitted from that tally, leaving two offsetting counting errors whose total impact remains unquantified.21 Furthermore, the researchers deliberately limited their queries to partial samples rather than downloading entire databases, and they reported no evidence showing whether unauthorized third parties had previously accessed or misused the exposed files.3
The broader pattern extends across the modern software landscape wherever rapid deployment obscures back-end architecture. Independent security evaluations across approximately 300,000 internet domains connected to Supabase have revealed readable database tables on 16,326 installations, demonstrating that the failure to establish row-level security affects many platforms that provision databases automatically.1 As natural language tools allow millions of non-technical creators to deploy web applications, the baseline requirement for database security remains unchanged: automated checks and generative assistants can suggest code, but human operators must verify that row-level permissions explicitly deny unauthorized reads before an application handles confidential records.21
The record
35 sentences trace to 3 sources.
- 1 Lovable Apps Exposed Passport Numbers and 200,000 Private Messages Without a Login See the source
- 2 Lovable apps expose passport details and more than 200K private messages See the source
- 3 AI app-maker Lovable apps expose personal details, researchers find thousands of cases See the source
