# AI-Built App Security Statistics 2026

> AI-built app security statistics for 2026: Lovable and Supabase app scans, an audit of 200 vibe-coded apps, and NVD and GitHub advisory counts, with a CSV.

Canonical: https://app369.com/statistics/ai-built-app-security-2026/
Last updated: 2026-11-08

What published scans, one academic audit and two vulnerability databases say about the security of apps built with AI tools such as Lovable and Claude Code, mostly Supabase-backed. Each study measured something different, so the numbers can't be combined.

Every figure below links to its source and was checked against that source on November 8, 2026. Where a number comes from a study a company paid for, the source says so.

## **Published Scans of AI-Built Apps**

Each study used its own sample and method, so compare them with care. The sample, date and publisher are in each source line.

| Metric | Value | Source |
| --- | --- | --- |
| **Lovable-built apps from Lovable's showcase with Supabase endpoints that had inadequate row-level security (homepages only, March 2025) **[Link to this finding](#aisec-palmer-lovable) | 170 of 1,645 apps (about 10.3%), 303 endpoints | [Matt Palmer, "Statement on CVE-2025-48757" (May 29, 2025; scan completed March 21, 2025; author then worked in developer relations at Replit)](https://mattpalmer.io/posts/statement-on-CVE-2025-48757/) |
| **CVE assigned for that finding **[Link to this finding](#aisec-palmer-cve) | CVE-2025-48757, CVSS 9.3 Critical (score supplied by MITRE; NVD analysis deferred; disputed by Lovable) | [NVD, CVE-2025-48757](https://nvd.nist.gov/vuln/detail/CVE-2025-48757) |
| **Sites using Supabase, which Symbiotic describes as vibe-coded, with at least one finding of any severity, including low-severity header issues (2026) **[Link to this finding](#aisec-symb-any) | 1,046 of 1,072 sites (98%) | [Symbiotic Security, "We scanned 1,072 vibe-coded apps" (June 2, 2026; a security vendor)](https://www.symbioticsec.ai/blog/we-scanned-1-072-vibe-coded-apps-98-had-security-flaws) |
| **Same scan: sites with a critical finding / a critical or high finding **[Link to this finding](#aisec-symb-critical) | 173 of 1,072 (16%) / 313 of 1,072 (29%) | [Symbiotic Security, "We scanned 1,072 vibe-coded apps" (June 2, 2026; a security vendor)](https://www.symbioticsec.ai/blog/we-scanned-1-072-vibe-coded-apps-98-had-security-flaws) |
| **Same scan: sites flagged as letting anyone delete records with only the public Supabase key **[Link to this finding](#aisec-symb-delete) | 172 sites | [Symbiotic Security, "We scanned 1,072 vibe-coded apps" (June 2, 2026; a security vendor)](https://www.symbioticsec.ai/blog/we-scanned-1-072-vibe-coded-apps-98-had-security-flaws) |
| **Same scan: sites with tables fully readable by anyone holding the public Supabase key **[Link to this finding](#aisec-symb-readable) | 39 sites | [Symbiotic Security, "We scanned 1,072 vibe-coded apps" (June 2, 2026; a security vendor)](https://www.symbioticsec.ai/blog/we-scanned-1-072-vibe-coded-apps-98-had-security-flaws) |
| **AI-built web apps using Supabase with a scanner-recorded row-level security or data-exposure finding (owner-submitted; scans December 17, 2025 to August 4, 2026; scanner output, not confirmed vulnerabilities) **[Link to this finding](#aisec-vas-rls) | 141 of 359 (39.3%) | [Vibe App Scanner, "What 1,215 VAS Scans Reported: 2026 Security Study" (August 5, 2026; a security vendor, owner-submitted apps)](https://vibeappscanner.com/blog/vibe-coded-app-security-report-2026) |
| **Publicly deployed apps built with Claude Code or Lovable that had at least one vulnerability (audit of 200 apps) **[Link to this finding](#aisec-arxiv-any) | 91.0% (1,186 vulnerabilities found) | ["Understanding the (In)Security of Vibe-Coded Applications", arXiv 2606.23130 v4 (revised September 14, 2026)](https://arxiv.org/abs/2606.23130) |
| **Same audit: share of the vulnerabilities rated Critical or High **[Link to this finding](#aisec-arxiv-severe) | 65.77% | ["Understanding the (In)Security of Vibe-Coded Applications", arXiv 2606.23130 v4 (revised September 14, 2026)](https://arxiv.org/abs/2606.23130) |
| **Publicly available vibe-coded apps analyzed, and what Escape.tech's scanner reported **[Link to this finding](#aisec-escape) | Over 5,600 apps; more than 2,000 vulnerabilities, 400+ exposed secrets and 175 instances of exposed personal data | [Escape.tech, methodology for its vibe-coded apps research (October 29, 2025; a security vendor)](https://escape.tech/blog/methodology-how-we-discovered-vulnerabilities-apps-built-with-vibe-coding/) |
| **Live apps built with Lovable, Bolt, v0, Replit or Base44 whose Supabase database answered, with at least one table readable without logging in (passive scan, August 12 to 14, 2026; scanner findings) **[Link to this finding](#aisec-reeve-supabase) | 2,096 of 3,680 (57%); 8,429 of the 30,998 apps scanned named a Supabase project | [Reeve, "The State of Vibe-Coded App Security 2026" (August 19, 2026; a vendor of an app scanner)](https://reeve.page/research/vibe-coded-app-security-2026) |
| **Same scan: apps with at least one finding of any severity, mostly missing browser security headers **[Link to this finding](#aisec-reeve-any) | 99% of 30,998 apps | [Reeve, "The State of Vibe-Coded App Security 2026" (August 19, 2026; a vendor of an app scanner)](https://reeve.page/research/vibe-coded-app-security-2026) |

## **Supabase Apps in General**

This scan covered every domain UpGuard found with signs of Supabase, not only apps built with AI tools, so it is kept apart from the scans above.

| Metric | Value | Source |
| --- | --- | --- |
| **Supabase databases exposing readable tables, found among about 300,000 domains with signs of Supabase use, whether or not the app was built with AI (UpGuard queried each for a "users" table) **[Link to this finding](#aisec-upguard) | 16,326 databases; over half had indicators of personal data (judged from table schemas, rows not read) | [UpGuard Research, "Everything Everywhere: Systemic Data Exposure in Supabase Apps" (page dated October 14, 2026, first published September 24, 2026; a security vendor)](https://www.upguard.com/blog/everything-everywhere-systemic-data-exposure-in-supabase-apps) |

## **One Incident, Measured**

A single app, included because the exposure was measured and published by the researchers who found it.

| Metric | Value | Source |
| --- | --- | --- |
| **One misconfigured Supabase database (Moltbook, an AI agent social network), exposed records **[Link to this finding](#aisec-moltbook) | 1.5 million API authentication tokens and 35,000 email addresses | [Wiz Research, February 2, 2026](https://www.wiz.io/blog/exposed-moltbook-database-reveals-millions-of-api-keys) |

## **Vulnerability Databases**

Georgia Tech's count covers open-source projects in general. For the Supabase and Firebase packages I checked, few library-level advisories turned up. NVD mentions of Supabase rose in 2026, and I did not investigate why. The exposures in the scans above come from how each app is configured.

| Metric | Value | Source |
| --- | --- | --- |
| **Vulnerabilities (CVEs) that Georgia Tech's Vibe Security Radar had traced to AI coding tools as of April 13, 2026, after scanning over 43,000 security advisories (the researchers say Claude Code and Copilot account for most) **[Link to this finding](#aisec-gt-radar) | 74 confirmed: 14 critical, 25 high | [Georgia Tech School of Cybersecurity and Privacy, "Bad Vibes: AI-Generated Code is Vulnerable, Researchers Warn" (April 13, 2026)](https://news.research.gatech.edu/2026/04/13/bad-vibes-ai-generated-code-vulnerable-researchers-warn) |
| **NVD records mentioning "Lovable" / "Bolt.new" **[Link to this finding](#aisec-nvd-lovable) | 1 (CVE-2025-48757) / 0 | [NVD API, keywordSearch="Lovable" (retrieved November 8, 2026)](https://services.nvd.nist.gov/rest/json/cves/2.0?keywordSearch=Lovable) |
| **NVD records mentioning "Supabase" (any product, not only Supabase itself) **[Link to this finding](#aisec-nvd-supabase) | 44 (2024: 3, 2025: 3, 2026: 38) | [NVD API, keywordSearch="Supabase" (retrieved November 8, 2026)](https://services.nvd.nist.gov/rest/json/cves/2.0?keywordSearch=Supabase) |
| **NVD records mentioning "Firebase" (any product, not only Firebase itself) **[Link to this finding](#aisec-nvd-firebase) | 38 (2024: 10, 2025: 4, 2026: 19; 5 earlier) | [NVD API, keywordSearch="Firebase" (retrieved November 8, 2026)](https://services.nvd.nist.gov/rest/json/cves/2.0?keywordSearch=Firebase) |
| **GitHub security advisories for the @supabase/supabase-js / @supabase/auth-js packages **[Link to this finding](#aisec-ghsa-supabase) | 0 / 1 (low severity) | [GitHub Advisory Database, affects=@supabase/supabase-js (retrieved November 8, 2026)](https://api.github.com/advisories?affects=%40supabase%2Fsupabase-js) |
| **GitHub security advisories for the firebase / firebase-tools npm packages **[Link to this finding](#aisec-ghsa-firebase) | 1 (medium) / 1 (low) | [GitHub Advisory Database, affects=firebase (retrieved November 8, 2026)](https://api.github.com/advisories?affects=firebase) |

## **What the Platforms Say**

| Metric | Value | Source |
| --- | --- | --- |
| **Supabase default for tables created in the dashboard **[Link to this finding](#aisec-supabase-default) | Row-level security is on by default; organization owners get weekly security emails | [Supabase, Security retro 2025 (January 7, 2026)](https://supabase.com/blog/supabase-security-2025-retro) |
| **What a table without row-level security allows, per Supabase **[Link to this finding](#aisec-supabase-norls) | If the table is in an exposed schema: readable and writable by any role with a grant on it | [Supabase docs, Row Level Security](https://supabase.com/docs/guides/database/postgres/row-level-security) |
| **What missing Firebase authentication and rules allow, per Google **[Link to this finding](#aisec-firebase-norules) | Anyone who guesses your project ID can steal, modify, or delete the data | [Firebase docs, Avoid insecure rules](https://firebase.google.com/docs/firestore/security/insecure-rules) |

## **Earlier Firebase Data**

Firebase addresses taken mainly from Android apps in 2021, before AI app builders. Included for context only.

| Metric | Value | Source |
| --- | --- | --- |
| **Share of tested Firebase databases found open in a 2021 test of addresses taken mainly from Android apps (not AI-built) **[Link to this finding](#aisec-avast-firebase) | 10.7% of about 19,300 databases tested (from about 180,300 addresses found) | [Gen Digital (Avast) research, September 2021](https://www.gendigital.com/blog/insights/research/research-shows-over-10-of-sampled-firebase-instances-open) |

## **What the numbers say **

The public scans I verified report AI-built apps with open or weakly protected data, but the results vary widely because each study tested something different: from 39 of 1,072 sites with fully readable tables in one scan, to 10.3% of 1,645 Lovable showcase homepages with inadequate row-level security in another, to 39.3% of 359 owner-submitted Supabase apps in one vendor's sample with a scanner-recorded row-level security or exposure finding, to 57% of 3,680 reachable Supabase-backed apps with a readable table in another. They can't be averaged into one number.

The scans mostly describe configuration problems rather than library bugs; that is my reading, not the finding of any one study. The four main Supabase and Firebase packages I checked have three GitHub security advisories between them, while the scans find tables without row-level security, rules that let anyone in, and secret keys in front-end code. Supabase's and Google's own docs say what happens then: a Supabase table in an exposed schema without row-level security is readable and writable by any role with a grant on it, and open Firebase rules let anyone who guesses the project ID read, change or delete data.

Defaults are improving. Supabase turned row-level security on by default for tables created in its dashboard. Its statement covers dashboard-created tables only.

## **Methodology **

I'm Simon Dziak and I run App369, a one-developer app studio in Miami. This page collects public data only. I did not scan, query or test anyone's app or database. On October 20, 2026 I searched for published scans and audits of AI-built apps, the National Vulnerability Database (NVD) API and the GitHub Advisory Database API, and Supabase's and Firebase's own documentation and blog posts. On November 8, 2026 I re-fetched every source and re-ran both API queries. The published figures were unchanged; the NVD keyword counts for Supabase and Firebase had risen, and the counts below are the new ones. The same day I added UpGuard's, Reeve's and Georgia Tech's figures, each read on its own published page.

A figure is included only if I could read it in the source itself, and each one links to that source. I left out numbers I saw only in summaries or secondhand posts, and one Android Firebase study whose date I could not confirm.

The NVD counts are keyword matches on record text, so "Supabase" and "Firebase" include products that merely mention them, not only flaws in Supabase or Firebase. I did not count "Cursor" because the word also matches database and mouse cursors.

### **Caveats**

- No study here is a random sample of all AI-built apps. Each study found its apps differently: web discovery, a public showcase, owner submissions or open-source repositories.
- Symbiotic Security, Vibe App Scanner, Escape.tech and Reeve are security vendors, and their counts are scanner findings of mixed severity, not confirmed vulnerabilities. UpGuard, also a security vendor, counted databases that returned readable tables.
- UpGuard found its domains through BuiltWith and the Chrome UX Report and checked each for a "users" table. Its sample is every Supabase site it could find, not only AI-built apps. Reeve scanned apps built with Lovable, Bolt, v0, Replit or Base44 from the outside, and reached the database for 3,680 of the 8,429 apps that named a Supabase project.
- Georgia Tech's radar counts a CVE only when the commit that introduced the bug carries an AI tool's signature, such as a co-author tag, so bugs with no signature, or whose markers were removed, are missed.
- The March 2025 Lovable scan was run by Matt Palmer, who, per the appendix of his statement, worked in developer relations at Replit, which also makes an AI app builder. It covered sites listed in Lovable's own showcase.
- Symbiotic's sites were found by an automated discovery system and checked with 80 automated security rules.
- Vibe App Scanner's apps were submitted by their owners, some domains were scanned more than once, and 1,007 of its 1,215 scans used a quick check set.
- The 200-app audit is an arXiv preprint (version 4, revised September 14, 2026), drawn from 9,041 open-source apps. Its figures may change in later versions.
- App369 sells fixes for AI-built apps, so I have a commercial interest in this topic.
- The Lovable scan is from March 2025. Lovable and Supabase have both added security checks since then.
- A public Supabase key in a site's JavaScript is normal and by design. The risk is a table or function that key can reach without the right rules.

[**Download the data (CSV) **](https://app369.com/data/ai-built-app-security-2026.csv)

You're welcome to use these figures. Please credit App369 and link to this page.

## **Cite this page **

To cite one number, link to its row with the “Link to this finding” link above. To cite the page, use:

> App369. “AI-Built App Security Statistics 2026.” https://app369.com/statistics/ai-built-app-security-2026/. Figures checked: November 8, 2026.

## **Frequently Asked Questions **

### **How many AI-built apps have security problems?**

It depends on the study. A 2025 scan of 1,645 Lovable showcase homepages found 10.3% with inadequate row-level security on their homepages, a 2026 vendor scan of 1,072 Supabase apps found 16% with a critical finding, and an academic audit of 200 deployed apps built with Claude Code or Lovable found 91.0% with at least one vulnerability of any severity.

### **Are Supabase and Firebase insecure?**

The four main packages I checked have three GitHub security advisories between them. The problems the scans find are configuration: tables without row-level security, Firebase rules that allow everyone, and secret keys in front-end code.

### **Is a public Supabase key in my JavaScript a vulnerability?**

No. The public (anon or publishable) key is meant to be in the browser. It becomes a problem when a table, view or function it can reach has no row-level security or a policy that lets everyone in.

### **Can I use these statistics?**

Yes. Download the CSV, credit App369 and link to this page. Each figure also links to its original source, which you should cite for the number itself.

## **Built Your App With AI? **

Run the free [Supabase checker](https://app369.com/tools/supabase-oct-30-checker/) on your live site, read the checklist for [Lovable](https://app369.com/blog/lovable-app-rescue-2026/) or [Bolt](https://app369.com/blog/bolt-app-rescue-2026/), or book a $450 Launch Check: within 3 business days you get each problem ranked by severity, and the $450 is credited if you go ahead with the fixes.

[](https://app369.com/fixed-price/ai-app-rescue/) [](https://app369.com/statistics/)