Vibe coding security incidents: 8 real cases and the check for each
14 November 2026· 10 min read · by Aaradhya
Is my app the next one? Every few weeks a new story lands about an AI-built app that leaked data, and you wonder if yours looks the same from the outside. This post covers eight real cases from 2025 and 2026. Each one comes with a check you can run yourself.
We link the original source for every case. Where the facts are disputed or unknown, we say so.
| Incident | Date | Root cause | What was exposed |
|---|---|---|---|
| Lovable, CVE-2025-48757 | Mar to May 2025 | Missing or weak row level security | API keys, personal data, payment details |
| Base44 auth bypass | Jul 2025 | Sign-up endpoints trusted a public app ID | Access to private apps |
| Replit agent and SaaStr | Jul 2025 | Agent could delete production data | Records on 1,206 executives and 1,196+ companies |
| Tea app | Jul 2025 | Reportedly a public storage bucket | About 72,000 images |
| Moltbook | Jan to Feb 2026 | Public key and no RLS policies | About 4.75 million records |
| Lovable-showcased EdTech app | Feb 2026 | 16 flaws, 6 critical | 18,697 user records |
| Lovable public-project exposure | Apr 2026 | Broken object-level authorization | Source code, credentials, chat history |
| DataTalks.Club and Terraform | Feb 2026 | Agent ran Terraform with auto-approve | Production database and snapshots |
1. Lovable and CVE-2025-48757
Matt Palmer found the problem on March 20, 2025. The CVE was published on May 29, 2025. A scan of 1,645 Lovable projects found 303 exposed endpoints across 170 projects, about 10.3%. The scan only looked at homepages, so it may undercount. Exposed data included API keys for Google Maps, Gemini and eBay, personal data and payment details. On one site, an outsider could write to the database and mark a payment as paid without Stripe.
Lovable disputes the CVE. It says customers are responsible for their own data, and the NVD record notes this. Palmer’s statement on the CVE gives his side.
Root cause: Supabase row level security (RLS) was missing or weak, so the public key could read and write tables.
Check your app. Run this in the Supabase SQL editor. Every row it returns is a table with RLS off.
select schemaname, tablename
from pg_tables
where schemaname = 'public' and rowsecurity = false;
A table with RLS on but a policy of using (true) is still open to everyone. For a fuller walkthrough, see Postgres row level security exposed.
2. Base44 auth bypass
Wiz reported this flaw to Wix on July 9, 2025. Wix fixed it within 24 hours, and Wiz published on July 29. Two undocumented endpoints, auth/register and auth/verify-otp, needed only the app_id. That ID isn’t secret. So anyone could create a verified account on a private app, even one set to SSO only.
Wix says it found no evidence of abuse. Wiz doesn’t say how many apps were affected.
Root cause: the sign-up flow treated a public ID as proof of permission.
Check your app. Call a protected route with no cookie and no token. You want a 401 or 403, not data.
curl -i https://YOUR_APP.example.com/api/YOUR_PROTECTED_ROUTE
Also look for sign-up or verify routes your AI tool added that you never asked for.
3. The Replit agent and SaaStr
In July 2025, Jason Lemkin was on a 12-day trial of Replit’s AI agent. On day 9, despite a code freeze, the agent deleted production records. By Lemkin’s account, that covered 1,206 executives and 1,196+ companies. Replit CEO Amjad Masad called it “unacceptable and should never be possible” on July 21, 2025. He promised automatic separation of dev and production databases.
The reports we read don’t settle whether the data was recovered. There is also an entry in the AI Incident Database.
Root cause: the agent had write access to the production database.
Check your app. List which database roles can delete or truncate. If your agent’s role is on the list, change that.
select grantee, table_name, privilege_type
from information_schema.role_table_grants
where table_schema = 'public'
and privilege_type in ('DELETE', 'TRUNCATE');
Better still, give the agent a dev database only. Keep production credentials out of its environment.
4. The Tea app
In July 2025, about 72,000 images from the Tea app were exposed. Roughly 13,000 were verification selfies and ID photos. Roughly 59,000 came from posts, comments and DMs. Tea says only users who signed up before February 2024 were affected.
The reported cause is a public Firebase storage bucket with no authentication. That came from a 4chan post. The Record says experts reportedly confirmed it. Tea hasn’t confirmed it officially.
Here is the part people get wrong. Tea is often cited as a vibe-coding breach. The evidence doesn’t support that label. Barracuda’s write-up says it isn’t known, and that the founder said two contract developers built the original app. We include it because the lesson still applies to you.
Root cause (reported): files sat in a storage bucket that anyone could read.
Check your app. Open a private window, log out, and paste the URL of a file that should be private. If it loads, it’s public. You can also test from the terminal.
curl -s -o /dev/null -w "%{http_code}\n" \
"https://YOUR_PROJECT.supabase.co/storage/v1/object/public/YOUR_BUCKET/YOUR_FILE"
A 200 on a private file is a leak. For Firebase, request https://firebasestorage.googleapis.com/v0/b/YOUR_BUCKET/o the same way.
5. Moltbook
Wiz reported this on January 31, 2026. The fix came on February 1, and Wiz published on February 2. About 1.5 million API auth tokens were exposed. So were 35,000 email addresses and 4,060 private messages between agents. That’s about 4.75 million records in total, and outsiders had write access. Wiz describes the app as vibe-coded.
Root cause: a Supabase key sat in client JavaScript, and the database had no RLS policies.
Check your app. Search your built files for secrets. This also flags Supabase JWTs, which start with eyJ.
grep -rEn "sk_live_|sk-[A-Za-z0-9]{20,}|sb_secret_|eyJ[A-Za-z0-9_-]{20,}\.eyJ" dist/
The anon key showing up is normal. A service_role key is not. Our post on the Supabase service_role key shows how to decode a hit and tell them apart.
Now test what the public key can do. This should return an error or an empty list, not rows.
curl "https://YOUR_PROJECT.supabase.co/rest/v1/YOUR_TABLE?select=*&limit=1" \
-H "apikey: YOUR_ANON_KEY" \
-H "Authorization: Bearer YOUR_ANON_KEY"
6. The Lovable-showcased EdTech app
The Register reported on February 27, 2026 that researcher Taimur Khan found 16 vulnerabilities in an app Lovable had showcased. Six were critical. A total of 18,697 user records were exposed.
Lovable says the project included code not written with Lovable. It also says the database wasn’t hosted by Lovable. So how much of this belongs to the tool and how much to the project is a fair question.
Root cause: the report lists 16 flaws, and the sources we read don’t name a single one.
Check your app. Pick one table with user data. Run the anon-key curl from case 5 against it. Then log in as user A and try to fetch user B’s record by ID. You should get a denial.
7. Lovable public-project exposure
The Register reported on April 20, 2026 that a free Lovable account could reach other users’ projects. That meant source code, credentials and AI chat history. The cause was a broken object-level authorization flaw. In plain terms, the server didn’t check who owned what it handed back.
The researcher says it affected projects created before November 2025. Lovable says only public projects were affected. Lovable first denied a breach, then apologised. No count of affected projects has been given.
Root cause: the server didn’t check ownership of each object.
Check your app. Treat anything you pasted into an AI chat or committed to a project as possibly readable. Search your code for keys, and rotate any you find. Our Lovable hacked checklist covers the steps.
grep -rEn "sk_live_|sk-[A-Za-z0-9]{20,}|sb_secret_|service_role" src/ .env* 2>/dev/null
8. Claude Code and Terraform at DataTalks.Club
On February 26, 2026, an AI coding agent ran a Terraform command with auto-approve while the right state file was missing. It deleted the production database, the network and the snapshots. One table held 1,943,200 rows, and about 2.5 years of data were involved. The team’s own snapshots were gone, but AWS support recovered one about 24 hours later.
Founder Alexey Grigorev wrote up the incident. He says he overrode the agent’s advice to keep the infrastructure separate. He calls it his own fault.
Root cause: an agent held the power to destroy production, and a human approved the setup.
Check your app. Do three things. Run terraform plan and read it before any apply. Turn on deletion protection. Then prove a restore works before you need one. If your database is on AWS RDS:
aws rds modify-db-instance \
--db-instance-identifier YOUR_DB \
--deletion-protection \
--apply-immediately
The pattern
Read the eight cases again and the same four causes keep showing up.
- A database open to the public key. Cases 1 and 5. RLS was off or empty.
- Secrets in browser code. Case 5, and credentials in case 7. Anything in the browser is public.
- Auth that trusts the client. Cases 2 and 7. A public ID or a missing owner check stood in for a real login.
- An agent with write access to production. Cases 3 and 8. The agent could delete what the team couldn’t afford to lose.
Case 4, the Tea app, is a fifth flavor: an open storage bucket. It isn’t clearly an AI story at all.
Some numbers put this in context, with a caveat. Veracode tests models on set coding tasks. It found that 45% of AI-generated code samples introduced a vulnerability in July 2025. Its October 2026 update found 38%. Those are lab tests, not a measure of live apps.
For live apps, Escape.tech scanned more than 5,600 public vibe-coded apps in October 2025. It reported 2,000+ vulnerabilities, 400+ exposed secrets and 175 instances of exposed personal data.
For a shorter list to start with, see the vibe coding security checklist. If an AI tool wrote your keys into code, see Claude Code generated API keys leak.
Or, we do it for you
Our free production audit runs 15 checks on your app and sends a scored report in 48 hours. If the report shows something that needs fixing, we do that as fixed-price Rescue work. We host AI-built apps on servers in Germany, with push-to-deploy from GitHub. Get a free production audit.
FAQ
Is vibe coding safe?
It can be, but the default setup often isn’t. The incidents above come from a few repeat mistakes, like missing RLS and keys in browser code. Run the checks in this post before you put real user data in the app.
How do I know if my Lovable app is secure?
Start with three checks. List tables with RLS off. Call your API with the anon key alone and see if it returns rows. Search your built files for secrets. Lovable says customers are responsible for their own data, so these checks are yours to run.
Is the Supabase anon key secret?
No. It’s meant to be public and ships in your frontend. What protects your data is RLS. A service_role key is different: it skips RLS, so it must never reach the browser.
Was the Tea app vibe coded?
The evidence doesn’t say so. Barracuda reports that it isn’t known, and that the founder said two contract developers built the original app. The reported cause, an open storage bucket, can happen in any app.
Do AI tools cause these breaches?
The sources don’t give one answer. In several cases a person or a platform setting shared the blame, and some platforms dispute the fault. The causes are old ones: open databases, exposed secrets and weak auth. AI tools can make them quick to create and easy to miss.
Get a free production audit
15-point check, scored report back in 48 hours, free either way.