research, November 2026
What we found in 1029 public apps built with AI
We read the code of 1029 public apps built with Lovable, Bolt, v0, Cursor, Claude Code and other AI tools, and 188 similar apps written before ChatGPT. 40% of the AI-built apps had at least one high or critical security problem.
app with a high or critical problem (409)no serious problem found (620)
40%
of AI-built apps had a high or critical problem
62%
of Lovable, Bolt and v0 apps did
50%
of apps that take payments did
15%
had their .env secrets file committed
The problems that came up most
Share of all 1029 AI-built apps with at least one case, from the automated checks. One noisy check is left out; the method page says which and why.
| Problem | Share of apps |
|---|---|
| Packages with a known security flaw | 41% |
| Raw HTML from users put on the page | 7% |
| User input built into SQL (Python) | 6% |
| Supabase Edge Function anyone can call | 5% |
| No limit on how often an endpoint can be called | 4% |
| API route with no login check | 4% |
| HTTPS certificate checks switched off | 4% |
Which apps had the most problems
Apps whose browser talks straight to the database fare worst; full-stack projects with a server in between fare best. Stacks, platforms and hosting
| Problem | Share of apps |
|---|---|
| Lovable, Bolt and v0 apps (201) | 62% |
| SaaS apps with payments, CRM or bots (100) | 57% |
| data-ml (40) | 55% |
| Python back ends (110) | 53% |
| web3 (22) | 46% |
| AI automations and agents (90) | 37% |
| ecommerce-content (41) | 34% |
| Infrastructure-heavy projects (40) | 33% |
| Mobile apps (71) | 25% |
| Full-stack Cursor and Claude Code projects (152) | 20% |
| browser-extensions (40) | 20% |
| Go, PHP, Ruby, Rust and Java back ends (80) | 19% |
| desktop-cli (42) | 12% |
AI fixed old mistakes and made new ones
Against apps created before ChatGPT, AI-built apps leak fewer keys and carry fewer outdated packages, but get access control and payments wrong far more often.
The full comparison| Measure | AI-built | Before ChatGPT |
|---|---|---|
| Any high or critical problem | 40%45% | |
| Leaked keys or secrets | 5%13% | |
| Strangers can read or change data | 8%1% | |
| Missing or broken login checks | 7%2% | |
| Payment holes | 3%0% | |
| Features anyone can use on your bill | 3%0% | |
| Key or password written in the code | 3%11% | |
| Outdated package with a known flaw | 53%59% |
Read by topic
AI-built vs pre-AI apps
1,029 AI-built apps against 188 similar apps created before ChatGPT: fewer leaked keys, more broken access control and payment holes.
Keys and secrets
15% of 1,029 AI-built apps had a .env secrets file committed to the repository. Where keys leak, and how to stop it.
Database access
8% of AI-built apps let strangers read or change data they should not. Row level security, open Firebase rules and database functions anyone can run.
Payments and webhooks
50% of AI-built apps that take payments had a serious problem; 9% had a hole in the payment flow itself.
AI features and costs
441 of 1,029 AI-built apps call an AI model themselves. How often those features can be used by anyone, on the owner's bill.
AI automations and agents
174 AI-built automation projects (n8n, Make and Zapier, AI agents, Telegram and Discord bots): 9% leaked a secret, 10% had an endpoint anyone could call.
Infrastructure and CI
223 AI-built projects ship Docker, Terraform, Kubernetes or CI files. 37% of them run containers as root.
Stacks, platforms and hosting
What 1,029 AI-built apps are made of: platforms, languages, the services they connect to and where they are hosted, and which groups had the most problems.
Method and limits
How we chose the repositories, what the automated checks can and cannot see, and the criticism this study deserves.
Earlier: our pilot study of 24 apps, where every finding was checked by hand.