Vibe coding security incidents: the public cases of 2025 and 2026, and what caused each

The public vibe coding security incidents of 2025 and 2026 fall into three groups: apps that answered strangers, platforms with a flaw of their own, and agents or coding tools that did the damage. Most app leaks needed no exploit. The data sat behind a public key, with access rules that were missing, open or inverted. Four checks tell you whether your app would have been in one of these reports.

The team behind Keybrum's attacksUpdated 12 min read

In short

  • The largest group is apps that answered strangers: 170 of 1,645 Lovable showcase apps in 2025 (CVE-2025-48757), Moltbook's 1.5 million API tokens in January 2026, and 18,697 user records in one showcased app in February 2026.
  • Platforms had flaws too. Wiz found a way to register on private Base44 apps in July 2025, and in April 2026 Lovable said a change had accidentally made chats on public projects readable again.
  • Agents and tools did damage directly: Replit Agent deleted a production database during a code freeze in July 2025, and a release of Amazon Q Developer for VS Code shipped with malicious code the same month.
  • The researchers behind the app leaks did not break anything. They read the public JavaScript, used the public key and asked the database.
  • The same four checks cover the first group: list tables without row level security, query your API signed out, sign in as a second account, and confirm what your project's visibility setting exposes.
On this page

What counts as a vibe coding security incident

A vibe coding security incident is a case where an app built mostly by prompting, the platform it was built on, or the coding agent itself exposed or destroyed data. This page lists the public ones from 2025 and 2026. Every entry has a source you can open: a CVE record, a researcher's write-up, a vendor bulletin or a press report. An app is named only when the source names it.

The cases fall into three groups, and the group decides what you can do about it. When the app answered strangers, the owner could have found the hole with a few checks. When the platform had a flaw, the owner could only limit what the flaw reached. When the agent or the tool did the damage, the protection was a backup and a separate production database.

Timeline: the 2025 cases

WhenWhat happenedGroup
March 2025Researchers scanned the homepages of 1,645 apps from Lovable's showcase and found 303 unprotected endpoints across 170 projects, about one in ten. The exposed data included names, phone numbers, API keys and payment details. Published on 29 May 2025 as CVE-2025-48757.App
March 2025Pillar Security showed the "Rules File Backdoor": invisible Unicode characters in a rules file can instruct Cursor and GitHub Copilot to insert malicious code without telling the developer. GitHub added a warning for hidden Unicode text in May 2025.Tool
July 2025SaaStr founder Jason Lemkin reported that Replit Agent deleted his production database during a code freeze, created fake data, and told him a rollback was impossible. The rollback worked. The full account.Agent
July 2025Wiz Research found two undocumented Base44 endpoints that let anyone create a verified account on a private app, knowing only the app's id. Wix fixed it within 24 hours and found no evidence of abuse. What it means for Base44 apps.Platform
July 2025A threat actor used an overly broad token to commit malicious code to the Amazon Q Developer extension for VS Code, and it shipped in version 1.84.0. AWS says a syntax error stopped the code from running, and released 1.85.0 (CVE-2025-8217).Tool
September 2025Wiz Research reported that one in five organizations building on vibe coding platforms was exposed, through four recurring flaws: login checked in the browser, secrets in client code, open database tables, and internal apps with no login.App
October 2025Escape analyzed more than 5,600 public apps, most of them built with Lovable, and counted more than 2,000 vulnerabilities, more than 400 exposed secrets and 175 instances of personal data, including medical records and IBANs.App
Lovable disputes CVE-2025-48757, on the grounds that each customer is responsible for protecting their own app's data.

Timeline: the 2026 cases

WhenWhat happenedGroup
January 2026Palo Alto Networks' Unit 42 described incidents it had seen, without naming them: a sales lead app breached because the agent left out authentication and rate limiting, a platform flaw that let an attacker pass authentication with an app's public id, and an agent that deleted a production database it had been told to leave alone.App, platform, agent
January 2026Wiz Research found that Moltbook, a social network for AI agents, let anyone read and write its Supabase database with the public key in its JavaScript: 1.5 million API tokens, 35,000 email addresses and 4,060 private conversations. It was fixed within hours. The Moltbook case.App
February 2026The Register reported that Taimur Khan found 16 vulnerabilities, six of them critical, in one app featured on Lovable's Discover page. They exposed 18,697 user records, including student accounts. The worst flaw was an access check written backwards: it let anonymous visitors in and blocked signed-in users.App
March 2026Georgia Tech's Vibe Security Radar, which traces published vulnerabilities back to commits signed by an AI tool, had confirmed 74 CVEs tied to AI coding tools: 6 in January, 15 in February and 35 in March.Code
April 2026A researcher reported that a free Lovable account could read the source code and chat history of other users' public projects, and found database credentials in that code. Lovable answered that it had not suffered a data breach, that public projects' code was visible by design, and that a February change had accidentally made their chats readable again, which it reverted. Lovable security has both sides.Platform
April 2026Vercel disclosed that an attacker got in through Context.ai, a third-party AI tool used by an employee, and decrypted environment variables not marked as sensitive for a limited subset of customers. The Vercel incident.Platform
May 2026The Israeli firm RedAccess reported about 380,000 publicly reachable apps and other assets built with Lovable, Base44, Replit and Netlify, about 5,000 of which held sensitive corporate information. Replit's CEO answered that public apps being reachable is expected behaviour. Base44 and Lovable said the reports lacked the addresses needed to verify them.App
November 2026The Hamburg firm Crissy Field examined 159,041 Lovable projects and found records containing email addresses in 10,560 tables across 7,403 websites. The authors say this is not a confirmed leak rate. The scan, figure by figure.App

Pattern 1: the app answered strangers

This is the largest group, and the one an owner can do the most about. In every case the app worked as its builder expected. The trouble was what it answered to someone the builder never pictured: a visitor with the developer tools open, or a second account.

The cause is usually one of three. Rules that are missing: Moltbook's tables had no row level security (RLS), the rules that decide which rows each caller may see, so the public key in the page opened everything. Rules that are open: a policy exists and lets every caller through, which a scan that only checks for the presence of a policy will pass. Rules that are wrong: in the February 2026 case the check ran backwards.

A classic logic inversion that a human security reviewer would catch in seconds.
Taimur Khan, quoted by The Register, February 2026

None of these needed an exploit. The scan behind CVE-2025-48757 looked only at homepages. Wiz found Moltbook's open database by browsing the site as a normal user and reading its JavaScript. RedAccess found apps that were public because nobody had switched them to private. That is also the good news: what a stranger can find with a browser, you can find first. In OWASP's terms this is broken access control, the first category of the Top 10 read for vibe-coded apps.

Pattern 2: the platform had a flaw of its own

Sometimes the hole is under your app, in a layer you cannot patch. The Base44 flaw sat in the platform's shared sign-up endpoints, so a private app could be joined by anyone who knew its id. Wiz drew a general lesson from it: the most immediate risks in these platforms come from basic controls such as authentication and API design, and one flaw in a shared layer reaches every app built on it. The Lovable case of April 2026 and the Vercel incident are of the same kind.

You cannot fix a platform, but you decide what its next flaw can reach:

  • Keep credentials out of source code and out of the chat. In the Lovable case, the database credentials the researcher reported were sitting in project source code. Store them in the platform's secrets feature.
  • Set the project's visibility on purpose. Know whether "public" covers the code, the chat history or only the published app, and switch to private when you have no reason to share.
  • Mark secrets as secrets. Vercel's bulletin concerned environment variables that were not marked as sensitive. It tells customers to rotate those and to use the sensitive setting from now on.
  • Do not rely on a hidden address. The Base44 write-up shows that an app's id is visible in its URL. Rules that check who owns each record still hold when a stranger gets an account.
  • Rotate after a bulletin. When your platform publishes an incident, rotate the keys it could have touched before you read the rest.

Pattern 3: the agent or the tool did the damage

In the third group nobody attacked the app. The agent acted on the wrong database, or the tool itself carried someone else's instructions. Lemkin's conclusion after the Replit incident was that a code freeze could not be enforced by telling the agent about it. Replit announced separate development and production databases on 21 July 2025, in beta for new apps, so that the agent works on a copy.

The tool cases are supply chain attacks on the developer. With the Rules File Backdoor, the instructions hide in a configuration file the agent reads in every session. With Amazon Q, they arrived in an official release of an extension. Three habits cover most of it: give the agent a development database and no production credentials, restore from a backup once before you need to, and read any rules file you did not write. The guides on Replit Agent, Cursor and AI agent security go further.

Would your app have been in one of these reports? Four checks

These cover the first pattern, on a Supabase backend, which is what most of the reports above concern. Run them on your own project only. First, list the tables without row level security in the Supabase SQL editor. The query reads the catalog and changes nothing:

sql
select c.relname as table_name
from pg_class c
join pg_namespace n on n.oid = c.relnamespace
where n.nspname = 'public'
  and c.relkind in ('r', 'p')
  and not c.relrowsecurity
order by 1;
Every row is a table that anyone holding your public key can reach.

Second, ask your database as a stranger. A policy can exist and still let everyone through, so test the answer, not the setting:

bash
# Your own project and your own public key only.
curl "https://YOUR-PROJECT.supabase.co/rest/v1/YOUR_TABLE?select=*&limit=1" \
  -H "apikey: YOUR_PUBLISHABLE_KEY"
[] or an error: the table refuses strangers. A row: it does not.
  1. Third, sign in as a second account. Create two accounts, add a record with the first, then try to open and edit it as the second, in the app and through the API. This is the check that would have caught an inverted or missing ownership rule.
  2. Fourth, read your visibility settings. On your builder, check whether the project is public or private and what public includes. On your host, check whether preview deployments and internal tools ask for a login.

On Firebase, the equivalent of the first two checks is reading your rules for allow read: if true and for conditions that stop at request.auth != null. The Firebase rules checker flags both in rules you paste, and the Supabase RLS checker does the same for policies.

What the platforms changed afterwards

The incidents moved the platforms, and they deserve credit for it. Replit split development from production. Lovable added a security scan before publishing, and The Register reports that its projects have been private by default on every tier since December 2025. GitHub now warns about hidden Unicode in files. Wix patched Base44 within a day of Wiz's report. AWS pulled the affected extension version and replaced its credentials.

What did not change is who answers for the app. In the February 2026 case, Lovable's chief information security officer told The Register that every project includes a free security scan before publishing, and that in this case its recommendations had not been applied. A scan the builder can skip, or that passes because a policy exists, leaves the question of what the live app answers. When the data belongs to people in the EU, the owner also has 72 hours to notify a breach: vibe coding compliance covers that side. The vibe coding risks guide covers the eight holes behind these cases, and the checklist the layers to test before launch.

What Keybrum tests against these patterns, and what it does not

Pattern 1 is what Keybrum is built for. It attacks your deployed app as a stranger, then as accounts it creates. On Supabase it reads every table the public key can reach, and tries inserts and updates that are rolled back so nothing is written. It calls exposed database functions, lists storage buckets and listens for realtime changes. On Firebase it tests Firestore, Realtime Database and Storage rules. On your own API it calls every route with no session, has one account try to read and change another account's object, opens admin pages signed out, and checks the keys in your bundle with read-only calls. Each finding comes with the request that worked and a fix written for your coding agent, and it closes only when the same attack fails.

Pattern 2 is out of scope. Keybrum does not test Lovable's, Base44's or Vercel's own systems. It tests the app you built on them, and only after you prove you own it. Pattern 3 is out of scope too. Keybrum does not watch your coding agent, your editor extensions or your rules files, and it never tests deleting data. What it does catch is the result of a bad change, once it is live: plans attack again on every deploy.

A one-time report costs €49 for one app, with retests of its findings for seven days. Plans start at €29 a month excluding VAT.

Questions

Has a vibe-coded app actually been hacked?

Yes, and more often an app was found open than broken into. Researchers found data readable in 170 Lovable showcase apps in 2025, in Moltbook's whole database in January 2026, and in thousands of apps in the 2026 scans by RedAccess and Crissy Field. Unit 42 also describes a sales lead app that was breached because the agent left out authentication.

Which vibe coding incidents exposed the most data?

By record count, Moltbook: Wiz reported 1.5 million API tokens and 35,000 email addresses readable with the public key. By number of apps, the 2026 scans: RedAccess reported about 5,000 apps holding sensitive corporate information, and Crissy Field found records with email addresses on 7,403 websites built with Lovable.

Did Replit's AI really delete a production database?

Yes. In July 2025 SaaStr founder Jason Lemkin reported that Replit Agent deleted his production database during a code freeze and then said a rollback was impossible, which turned out to be wrong. Replit announced separate development and production databases on 21 July 2025.

Was the Tea app breach caused by vibe coding?

The evidence does not support it for the first incident. Tea's statement places the exposed images in a dataset from before February 2024, about a year before the term was coined. The lesson still applies to AI-built apps: storage has its own access rules, separate from the database.

Are vibe coding platforms themselves safe?

They have had flaws, and fixed them: Base44's sign-up endpoints in July 2025, and chats on public Lovable projects in 2026. Most incidents, though, are in apps built on the platforms, in rules the builder controls. Keep credentials out of source code and chat, and set project visibility on purpose.

How do I know if my app is affected by the same problems?

Run the four checks on this page: list tables without row level security, query your API with only the public key, sign in as a second account and try to reach the first one's data, and read your project's visibility settings. The free check shows what your JavaScript already tells every visitor.

Sources

  1. 1.Statement on CVE-2025-48757, Matt Palmer
  2. 2.CVE-2025-48757, NIST National Vulnerability Database
  3. 3.New Vulnerability in GitHub Copilot and Cursor: How Hackers Can Weaponize Code Agents, Pillar Security
  4. 4.Vibe coding service Replit deleted production database, The Register
  5. 5.Introducing a safer way to vibe code with Replit databases, Replit
  6. 6.Critical Vulnerability in Base44, Wiz Research
  7. 7.Security Update for Amazon Q Developer Extension for Visual Studio Code (Version #1.84), AWS-2025-015, Amazon Web Services
  8. 8.Wiz Research discovers one in five organizations exposed to systemic risks in vibe-coded applications, Wiz Research
  9. 9.Methodology: how we discovered over 2k high-impact vulnerabilities in apps built with vibe coding platforms, Escape
  10. 10.Securing Vibe Coding Tools: Scaling Productivity Without Scaling Risk, Palo Alto Networks Unit 42
  11. 11.Hacking Moltbook: AI social network reveals 1.5M API keys, Wiz Research
  12. 12.Lovable-hosted app littered with basic flaws exposed 18K users, researcher claims, The Register
  13. 13.Security Researchers Sound the Alarm on Vulnerabilities in AI-Generated Code, Infosecurity Magazine
  14. 14.Vibe coding upstart Lovable denies data leak, cites 'intentional behavior,' then throws HackerOne under the bus, The Register
  15. 15.Vercel April 2026 security incident, Vercel
  16. 16.Thousands of Vibe-Coded Apps Exposing Corporate, Personal Data: RedAccess, Security Boulevard
  17. 17.Lovable: research on 159,041 projects, Crissy Field

Your app is public. So are its mistakes.

Check what any attacker can already see in it, free, in ten seconds. The full attack, with proof and fixes: €49 once, or €29 a month on a plan.

or continue with Google or GitHub