Production Ready
What goes wrong with AI-built apps
A dated list of real incidents, what caused each one, and the fix.
By Joe Minock · Last reviewed
The same few problems keep showing up: data open to anyone, secret keys in the browser, logins that don't really check, agents with too much power and no backup, and runaway bills. Here are the public incidents behind them, oldest first, with sources. We add new incidents as they're reported.
Key numbers
170 of 1,645
apps on Lovable's showcase let anyone read user data ()
Over 400
exposed secrets in 5,600 public apps built with AI tools ()
1.5 million
API tokens exposed by one database key in browser code ()
$600,000
in AI credits used on one stolen key with no spending cap (disclosed )
About 5,000
of about 380,000 public apps found leaking data ()
45%
of the time, AI models chose the insecure way on coding tasks (Veracode)
The incidents
Oldest first. Select a date to link to that row. Scroll the table sideways to see every column.
| Date | Platform | What happened | Cause | Platform bug or builder mistake | Fix | Source |
|---|---|---|---|---|---|---|
| to | Lovable | Lovable apps open to anyone A researcher scanned 1,645 apps on Lovable's showcase and found 170 letting anyone read user data, including names, emails, financial details and secret keys. | Missing or weak row-level security in Supabase. | Unclear | Row-level security on every table, with each policy checked. | CVE-2025-48757, reported by Semafor |
| Base44 | Base44 private apps open to sign-up Anyone with an app's ID could register on a private Base44 app and get past login. It was fixed within 24 hours. | A flaw in Base44. | Platform bug | Test that private really means private. | Wiz | |
| Replit | Replit's agent deletes a live database During a code freeze, Replit's agent deleted SaaStr's production database, with records for about 1,200 executives and 1,200 companies. | The agent could reach live data. | Unclear | Keep test and live data separate, never give an agent access to live data, and keep backups you've tested. | Fortune | |
| Gemini CLI | Gemini CLI destroys a user's files The agent destroyed a user's files while reorganizing them. | It misread a failed command. | Unclear | Agents ask before running commands, and work on a copy. | AI Incident Database | |
| Replit | Surprise agent bills One user said Replit's agent charged about $1,000 in a week, against about $200 a month before. | Replit changed its agent pricing. | Unclear | Spending limits and usage alerts. | The Register | |
| Several AI tools | Thousands of flaws across 5,600 apps Escape scanned 5,600 public apps built with AI tools: more than 2,000 vulnerabilities, over 400 exposed secrets and 175 cases of exposed personal data, including medical records. | Not stated. | Unclear | A security review before launch, not after. | Escape | |
| Google Antigravity | Google's Antigravity wipes a drive Asked to clear a project cache, the agent wiped a user's whole drive. | It was set to run commands without asking. | Unclear | Agents ask before anything destructive. | The Register | |
| Supabase | Moltbook exposes 1.5 million API tokens Moltbook exposed 1.5 million API tokens, about 35,000 emails and thousands of private messages. | Its database key sat in browser code with row-level security off. | Builder mistake | Keys on the server, and row-level security on. | Wiz | |
| Lovable | A login check that worked backward An app on Lovable's own showcase let anonymous visitors in and blocked signed-in users, exposing about 18,700 user records. | The app's login check worked backward. | Builder mistake | A person reads the login logic. Scanners miss this. | The Next Web | |
| Claude Code | A Stripe key in browser code A founder wrote that attackers used the live Stripe key in his app to charge 175 customers $500 each. Stripe reversed the charges but kept $2,500 in fees. | His live Stripe secret key was in browser code. | Builder mistake | Secret keys only on the server, and rotate any that were ever exposed. | Awesome Agents | |
| Several AI coding tools | Public vulnerabilities traced to AI coding tools Georgia Tech researchers traced 35 public vulnerabilities in a single month to AI coding tools, up from 6 in January. | Not stated. | Unclear | Review AI-written code like any other code. | Cloud Security Alliance | |
| (disclosed ) | Not named | A $600,000 AI bill An attacker got the model's API key and used about $600,000 in AI credits over three weeks. Nobody noticed. | A bug in an AI-built app turned off the login on a public server, and the key had no spending cap. | Builder mistake | Spending caps and alerts on every key. | METR |
| Lovable | Lovable leaves projects readable for 48 days Any free account could read other users' source code, AI chats and database keys on some public projects, for 48 days. | A flaw in Lovable. | Platform bug | Own your code and hosting, and rotate keys after any platform incident. | The Register | |
| Lovable, Base44, Replit and Netlify | About 5,000 apps leaking data A scan of about 380,000 public apps found about 5,000 leaking corporate or personal data, including medical and financial records. | Apps are often public by default. | Unclear | Check whether yours is public. | Axios |
What the research says
- Insecure choices Veracode tested more than 100 AI models on coding tasks. They chose the insecure way 45% of the time.
- More issues per change CodeRabbit found pull requests co-written with AI had about 1.7 times more issues than human-written ones.
- Feeling faster isn't being faster In a METR study, experienced developers were 19% slower with AI tools on their own projects, while believing they were 20% faster.
- "Almost right" In Stack Overflow's 2025 survey, 66% of developers said their top frustration was AI answers that are almost right, but not quite.
What to do about it
Run our ten-point security checklist today. If anything fails, or you'd like a second set of eyes, start with a Production Ready review.
Changelog
: first published.
Common Questions
Frequently asked questions
Is vibe coding safe?
It can be. The tools are fast and useful. The risk is shipping what they build without the checks a production app needs: data rules, keys on the server, real logins, backups and spending limits.
Are these platform bugs or builder mistakes?
Both. Lovable's 48-day exposure in April 2026 and Base44's login flaw in July 2025 were platform bugs. Most leaks, like tables with row-level security off and keys in browser code, came from how the app itself was built.
How do I know if my app is affected?
Run our ten-point security checklist. If you find tables without row-level security, keys in browser code, or private data you can reach without signing in, fix those first.
Proof
How we built Deliberate with Claude Code
Deliberate is our own Mac app, built by a three-person team with Claude Code. We built the app, its registration and billing site, and integrations with Gmail, Outlook, Slack, Asana, Zoom and Granola. It runs on your Mac, never saves the full text of your email, and sends nothing without your yes.
Meet DeliberateWorried your app is on this list?
Send us the link. We'll tell you what we'd fix first.