Ground Truth.
AI, checked against the source.

News · 2026-11-15

Researchers find sensitive records exposed in Lovable-built apps

Crissy Field researchers say they found private messages, identity-document information and other sensitive records accessible through some apps built with Lovable. The finding puts the security of AI-generated application defaults and database permissions under scrutiny, but does not establish a platform-wide breach or a representative leak rate.

Key facts

The researchers used Common Crawl to discover candidate sites, then investigated database responses separately. They found 34,545 projects using Supabase, and selected sites where database records were returned on the initial page load without login or other interaction. Publicly visible database records are not necessarily private: an app may intentionally display contact information, for example. The subsequent email-address test narrows the search but cannot reliably distinguish private personal information from public or test data. It also misses sensitive records without email addresses.

The concrete examples are more telling than an extrapolated percentage. The report describes more than 80,000 forum profiles and more than 200,000 chat messages, over 1,000 records with identity-document details, and other customer and business records. The researchers say they retrieved parts of datasets rather than complete copies. They could not determine whether anyone else accessed or misused the information. Their investigation establishes what they report observing, not the complete exposure period or a count of victims.

The technical issue is the authorization rule behind an application, not simply the presence of a key in a web page. Supabase explains in its database security guidance that a client-facing publishable key can be used in a browser when database grants and row-level security policies correctly restrict what each user can see. Think of the key as identifying a public reception desk: the important question is which filing cabinets the receptionist is allowed to open. A browser-visible key alone is not evidence of a stolen master key.

Lovable’s head of security Igor Andriushchenko said in the public discussion: “The root cause in these apps was insufficient Row-Level Security.” He also described scanning and remediation options. Lovable’s separate security post says users can review findings, use an agent to fix them, opt into automatic fixes, and configure a publication block for critical issues. The public record does not show which warnings the affected apps received or which controls were available when each was published.

The strongest counterpoint is that some anonymous database access is intentional, and app owners control their own access rules. That is true, but it does not explain away the specific sensitive examples the researchers describe. The meaningful product question is whether a builder with little database experience can understand and enforce the intended boundary before publishing. Commenters also disputed how much data researchers should retrieve to demonstrate the issue; the report’s findings should not be confused with a legal ruling on that question.

For app owners, the lesson is to inventory exposed tables, test access as an anonymous and ordinary user, and verify that ownership rules actually prevent cross-user reads. The prompt injection explainer covers a different AI security boundary: instructions hidden in content. Here the immediate problem is ordinary database authorization made consequential by the speed and accessibility of app generation.

The difference between finding a page and finding its data is important to the investigation. The researcher said Common Crawl was used to discover candidate sites; the private records were not stored in the crawl itself. The additional requests were made to app-linked services. That distinction keeps the discovery dataset from being mistaken for the source of the sensitive records and helps explain why the finding concerns an active application’s permissions.

An application can have a public catalog and private customer accounts in the same database. A test that confirms the catalog is readable therefore does not establish that account records should be readable. Checking the boundary requires representing several roles: an anonymous visitor, a signed-in user, and another signed-in user who does not own the record. The intended result depends on the application, so a broad count alone cannot decide which rows were improperly exposed.

The company’s scanning description also leaves a consequential chronology open. The available evidence does not show whether the current scanner existed when each affected app was published, whether it flagged the relevant configuration, or whether a user saw and overrode a warning. Those unanswered questions are important for assigning responsibility, but they do not erase the researchers’ reported examples. The report does not supply access logs establishing a previously unknown attacker or the duration of exposure.


Primary source, verified: read the paper →

Key questions

Did the researchers prove 30% of all Lovable apps leaked data?

No. The 30% figure counts sites with email-containing records within a selected group of 24,664 sites, and email addresses can be public or test data.

Was Lovable’s central platform breached?

The report describes app-level access-control exposures, not a demonstrated compromise of Lovable’s central service.

Did attackers misuse the exposed records?

The researchers say they could not establish whether the data was abused.
Cite this

APA

Ground Truth. (2026, November 15). Researchers find sensitive records exposed in Lovable-built apps. Ground Truth. https://groundtruth.day/news/lovable-apps-exposed-sensitive-records-researchers-report.html

BibTeX

@misc{groundtruth:lovable-apps-exposed-sensitive-records-researchers-report,
  title  = {Researchers find sensitive records exposed in Lovable-built apps},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/lovable-apps-exposed-sensitive-records-researchers-report.html}
}

Topics: cybersecurity · ai-security · access-control · vibe-coding

Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.