Most writing about vibe coding risk is hypothetical. This page is not. It lists the incidents involving AI app builders and AI coding agents that we could confirm from a primary or reputable source, newest first, with the date, what happened, the root cause in one line and the check that would have prevented it. Where a link to vibe coding is disputed, we say so.
The pattern is consistent. Very few of these incidents needed a clever attacker. They came from open database rules, secrets left in readable places, agents given more permission than the job needed, and bugs on the platform itself. If you are new to the term, start with what vibe coding is and where it breaks. This is a living article: we review it and add incidents as new ones are confirmed.
- Missing or wrong Row Level Security is the most repeated cause: Moltbook, a Lovable-hosted education app and the disputed CVE-2025-48757 all trace back to it.
- Apiiro found 1,379 distinct secrets in public Claude Code settings files in November 2026, and 379 of them were database passwords.
- Agents have deleted production databases, a home directory and a whole drive. Each time, the agent could run destructive commands without a human approving them.
- Some incidents were the platform's fault, such as Base44's 2025 auth bypass and Lovable's 2026 public-project exposure. You still own the response.
- A one-week check of RLS, secrets, agent permissions and backups covers every pattern on this list.
Which vibe-coding security incidents have actually happened?
Thirteen confirmed or widely reported events between May 2025 and November 2026, listed below. Two of them are disputed as vibe-coding incidents, and one is a large scan rather than a single breach. Each row links to the source we opened.
| Date | Tool | What happened | Lesson |
|---|---|---|---|
| Nov 2026 | Claude Code | Apiiro found 1,379 secrets in public .claude/settings.local.json files. Root cause: approved commands, passwords included, saved in plain text and committed. | Ignore .claude/ local files in Git, scan history and rotate anything found. |
| Jul–Aug 2026 | Codex (GPT-5.6 Sol) | Users reported real files deleted. Root cause: a temp-folder cleanup hit the real $HOME, mostly in full-access mode. Safeguards shipped in August. | Keep the sandbox on. Never give an agent full access to a machine with unbacked-up data. |
| Feb–Apr 2026 | Lovable platform | Chat history and source code of public projects were readable from 3 February to 20 April. Root cause: a backend permission regression; early reports were closed in triage. | Never paste secrets into a builder's chat. Rotate any that were. |
| 26 Feb 2026 | Claude Code and Terraform | DataTalks.Club's production database and network were destroyed, with its automated snapshots. Root cause: an old state file plus an agent-run terraform destroy. | Deletion protection on, backups outside the stack, humans run destructive commands. |
| Feb 2026 | Lovable-hosted app | A researcher reported 16 flaws in an education app, exposing 18,697 user records. Root cause: AI-written auth logic that blocked signed-in users and let anonymous ones through. | Test every function signed out and as a second user. |
| 31 Jan 2026 | Moltbook (vibe-coded) | Wiz found the Supabase database open to read and write. Root cause: a key in client JavaScript with no Row Level Security. | Enable and test RLS on every exposed table. |
| Dec 2025 | Claude Code | A user reported their Mac home directory deleted. Root cause, from Claude's own log analysis: ~/ included in an rm -rf command. | Deny rm -rf outside the project. Keep a current backup. |
| Late Nov 2025 | Google Antigravity | A user's whole D drive was wiped while clearing a project cache. Root cause: Turbo mode ran the command without approval. | Turn off auto-run for shell commands on your own machine. |
| 29 Oct 2025 | Lovable, Base44, Create.xyz and others | Escape scanned 5,600+ public apps and found 2,000+ vulnerabilities and 400+ exposed secrets. Root cause: default and missing access rules. | Assume your live app has been scanned. Scan it yourself first. |
| Jul 2025 | Base44 platform | Wiz found an auth bypass into private apps using only a public app_id. Fixed within 24 hours; no evidence of abuse. | Check sign-up logs after any platform disclosure. |
| Jul 2025 | Replit agent | SaaStr's production database was wiped during a code freeze. Root cause: the agent could change live systems. | Separate development and production databases. |
| Jul 2025 | Tea app (disputed) | About 72,000 images leaked from legacy storage. Simon Willison is confident vibe coding was not the cause. | Old storage buckets need the same review as new code. |
| May 2025 | Lovable (disputed) | CVE-2025-48757: weak RLS in generated apps. Lovable disputes it, saying each customer is responsible for their own app's data. | Your data rules are yours, whatever the platform says. |
Two things stand out. Agent incidents, where a tool deletes real data, have recurred every few months since July 2025. And a vendor saying "the customer is responsible" does not change what UK GDPR expects from the business that holds the data.
Why do vibe-coded apps keep leaving their database open?
Because the front end talks straight to the database with a public key, and the only thing deciding who sees which rows is Row Level Security. When the AI skips it, writes it backwards, or the builder never asks, the whole table is public.
Moltbook is the clearest case. Its founder said publicly that he did not write a single line of the code. Wiz found a Supabase key in the client-side JavaScript and no RLS policies behind it, which gave anyone read and write access to production. Wiz's own summary: "without RLS policies, this key grants full database access". The team fixed it in about three hours after first contact.
- 1.5 million API authentication tokens
- 35,000 email addresses
- 4,060 private direct-message conversations
- About 3 hours from first contact to all tables secured
Source: Wiz Research, Exposed Moltbook database, February 2026
The February 2026 education app shows a subtler version. RLS was not the only gap. The researcher said an AI-generated access check had its logic inverted, so it blocked signed-in users and allowed anonymous ones. That passes a happy-path demo, because the owner is usually testing while signed in. Lovable said the project included non-Lovable code and that the database was not hosted by Lovable, which is a fair point about scope but not about risk.
CVE-2025-48757 is the disputed one. The researchers behind it scanned 1,645 Lovable showcase projects and reported 170 with inadequate RLS. NVD lists the record as disputed, because Lovable says each customer is responsible for protecting their app's data. Both things can be true: the platform made the mistake easy, and the app owner still had to fix it.
- 5,600+ publicly available vibe-coded apps analysed
- 2,000+ vulnerabilities identified
- 400+ exposed secrets
- 175 instances of exposed personal data, including medical records and IBANs
Source: Escape, How we discovered vulnerabilities in apps built with vibe coding, 29 October 2025
About 4,000 of the apps Escape found were on Lovable, 449 on Create.xyz and 159 on Base44. Our Supabase and Lovable security checklist shows how to test RLS with two accounts and a signed-out browser.
How do secrets leak from AI-built apps and coding agents?
Mostly through files and chats that nobody thinks of as code. Keys end up in front-end bundles, in builder chat history, and now in coding agent settings files that get committed to public repositories.
Apiiro's November 2026 research is the largest example. When a Claude Code user picks "Yes, and don't ask again", the approved command can be saved to .claude/settings.local.json, including any password typed into it. Apiiro searched public GitHub and found 104,185 copies of the file in 103,617 repositories. Of those, 1,122 held at least one secret, adding up to 1,379 distinct secrets.
Base: 1,379 distinct secrets in 1,122 public .claude/settings.local.json files
| Secret type | Count |
|---|---|
| Database passwords | |
| SSH logins for public servers | |
| GitHub, GitLab and npm tokens | |
| AWS keys |
Source: The Hacker News, reporting Apiiro research, 7 November 2026
Two details make this worse. Standard scanners missed a lot: Gitleaks flagged 564 of the 1,122 files, TruffleHog flagged 643, and 215 were missed by both, mostly passwords typed inside commands. And the files stayed up: in half of the affected repositories the secret had been on the default branch for at least 217 days. Claude Code adds the file to the user's global git excludes, but that does not help once a file is already committed.
Lovable's April 2026 incident is the chat-history version of the same problem. For public projects, anyone with the link could read the build conversation from 3 February to 20 April. Lovable says private projects and Lovable Cloud were not affected. But if you ever pasted an API key into a public project's chat, treat it as exposed and rotate it. The Claude Code to production checklist covers where agent files and keys should live.
Can an AI coding agent delete your production database?
Yes. It has happened at least twice to production databases and several more times to local files. In every case, the agent was allowed to run a destructive command without a human approving that specific command.
In July 2025, Replit's agent wiped SaaStr's production database during a declared code freeze. Fortune reported data on more than 1,200 executives and over 1,190 companies. Jason Lemkin recovered the data. Replit's CEO called it "unacceptable and should never be possible" and introduced automatic separation of development and production databases.
In February 2026, Alexey Grigorev of DataTalks.Club lost about 2.5 years of course submissions. Working from a new computer without the Terraform state file, he let Claude Code run terraform destroy. An older state file pointed it at production. The database, network and automated snapshots went. AWS support restored a hidden snapshot 24 hours later. His own verdict: "This incident was my fault." He now runs every destructive command himself and restores a backup automatically every night to prove it works.
The local-machine cases follow the same shape. Google Antigravity in Turbo mode wiped a user's D drive while clearing a cache. A Claude Code user reported an rm -rf that included ~/. In July 2026, OpenAI's head of Codex said GPT-5.6 Sol sometimes overrode $HOME for a temp folder and then deleted the real one, more often in full-access mode without the sandbox. These are user reports or vendor statements, not audits, but the lesson holds.
If your team uses these tools daily, our Claude Code training covers permission settings and hooks that block destructive commands. The Codex pre-launch checklist covers the same ground for Codex.
What about security incidents on the platform itself?
Some incidents were nobody's fault but the vendor's, and you could not have prevented them. You can still limit the damage and you still own the response to your users.
Base44's July 2025 flaw let anyone create a verified account on a private app, bypassing single sign-on, using only the app_id visible in its URL. Wiz reported it on 9 July and Wix fixed it within 24 hours, with no evidence of abuse. Lovable's 2026 exposure was a backend regression: researchers reported it from 22 February, the reports were closed because triage documentation described the access as intended, and it was fixed within two hours of a public report on 20 April.
Tea is the counter-example. The July 2025 leak of about 72,000 images was widely blamed on vibe coding. Tea's statement points to legacy storage holding data from before February 2024, and Simon Willison is confident vibe coding was not to blame.
For platform incidents, three things help. Keep secrets out of builder chats and public projects. Check sign-up and access logs when a disclosure appears. And know which tables hold personal data, so you can decide quickly whether to report. The ICO's personal data breach guide requires a report within 72 hours of becoming aware when people's rights and freedoms are likely to be at risk, whoever caused it. Our guide to the first 72 hours after a vibe-coded app is hacked walks through that clock.
Which control would have prevented each type of incident?
Five controls cover almost everything above. No single one covers all four patterns, which is why a scan alone is never a sign-off.
| Incident pattern | RLS with a two-account test | Secret scanning and rotation | Agent sandbox and approvals | Separate dev and prod | Tested off-platform backups |
|---|---|---|---|---|---|
| Open database (Moltbook, education app, CVE-2025-48757) | Yes | Partial, catches keys in the front end | No | No | Partial, restores altered data |
| Leaked secrets (Apiiro, Lovable chats) | No, passwords bypass RLS | Yes | No | Partial, limits a dev key's reach | No |
| Agent deletes data (Replit, DataTalks.Club, Antigravity, Codex) | No | No | Yes | Yes | Yes |
| Platform-side flaw (Base44, Lovable 2026) | Partial, limits what an intruder reads | Partial, rotate what was exposed | No | No | Partial |
Sources: Wiz, Moltbook; Apiiro via The Hacker News; Alexey Grigorev, DataTalks.Club post-mortem; Lovable, April 2026 incident. Mapping by Circlemount.
If your app talks to Supabase from the browser, the first column is non-negotiable. If anyone uses a coding agent with shell access, so are the third and fifth.
What should you check this week?
Seven checks, about a day of work for a small app. They are ordered so the riskiest gaps close first.
- Open the Supabase Security Advisor and fix every table flagged without Row Level Security.
- Sign in as two test users and try to read each other's rows, then repeat signed out.
- Search the repository and its history for
.claude/,.envand secret or service_role keys. Rotate anything found before you delete it. - Add
.claude/settings.local.jsonto the project.gitignoreand turn on secret scanning in CI. - Turn off auto-run, Turbo and full-access modes. Deny
rm -rf,terraform destroyand database drop commands without approval. - Point agents at a development database, never production, and turn on deletion protection.
- Restore your latest backup to a scratch database and run one real query against it.
If step 2 or step 3 finds something, assume it may already have been seen. Rotate first, then work out what was reachable and for how long, then decide whether the ICO needs to hear about it.
So, is vibe coding safe for a business app?
It can be, if someone checks the parts the AI does not think about. Every incident here had a known, boring fix. Shipping without those checks is the real risk.
We use these tools ourselves. The Study Room's Portal, Dentorb and ORBQ were built with AI-assisted development using Antigravity, Claude Code, Codex and Lovable, and the same checks above are part of how we ship. See The Study Room, Dentorb and ORBQ. If you would rather have someone run them for you, our vibe-coded app security audit works through every pattern on this page, and our team can fix a vibe-coded app when the audit finds more than a few gaps.
Frequently asked questions
What are the biggest vibe coding security risks?
Confirmed incidents point to four: databases without working Row Level Security, secrets left in front-end code, chats or agent settings files, coding agents allowed to run destructive commands, and flaws on the platform itself. The first three are within your control and can be checked in about a day for a small app.
Has an AI coding agent really deleted a production database?
Yes. In July 2025 Replit's agent wiped SaaStr's production database during a code freeze, and in February 2026 Claude Code ran terraform destroy against DataTalks.Club's production setup. Both were recovered. In each case the agent could run destructive commands on production without a human approving that command.
Was there a Lovable security breach?
Lovable confirmed that from 3 February to 20 April 2026 the chat history and source code of public projects could be read by other users. It fixed the issue within two hours of a public report. Private projects and Lovable Cloud were not affected. Separately, CVE-2025-48757 about weak RLS is disputed by Lovable.
Did vibe coding cause the Tea app leak?
Probably not. Tea said the leaked images came from a legacy storage system holding data from before February 2024, and Simon Willison said he was confident vibe coding was not to blame. It is often cited as a vibe-coding incident, but the evidence points to an old, unprotected storage bucket.
How do I stop Claude Code leaking secrets into GitHub?
Add .claude/settings.local.json to the project gitignore, never type passwords directly into approved commands, and reference secrets as environment variables instead. Turn on secret scanning in CI that covers git history. If the file was ever committed, rotate every secret it contained before removing it from the repository.
Is vibe coding safe for an app with customer data?
It can be, but only after a human has checked the database rules, keys, agent permissions and backups. Every incident in this timeline had a known fix. Under UK GDPR you remain responsible for the data whatever tool wrote the code, including reporting a notifiable breach to the ICO within 72 hours.
Sources
Checked on 20 November 2026. Product features and figures change quickly, so check the current source before you rely on a detail.
- The Hacker News: Claude Code's don't-ask-again file exposes secrets (Apiiro research, 7 November 2026)
- BankInfoSecurity: OpenAI warns GPT-5.6 file deletions stem from full-access mode (17 July 2026)
- Lovable: Our response to the April 2026 incident (22 April 2026)
- Alexey Grigorev: How I dropped our production database (6 March 2026)
- The Register: Lovable-hosted app exposed 18K users, researcher claims (27 February 2026)
- Wiz Research: Exposed Moltbook database reveals millions of API keys (2 February 2026)
- Escape: How we discovered vulnerabilities in apps built with vibe coding (29 October 2025)
- Wiz Research: Critical vulnerability in Base44 (29 July 2025)
- Fortune: Replit AI coding tool wiped a database (23 July 2025)
- NVD: CVE-2025-48757 (disputed)
Check your app before it joins this list
Our vibe-coded app security audit tests your Row Level Security, keys, agent setup, payment flows and backups against every pattern in this timeline, and tells you plainly what to fix first. The initial app assessment is free, project work starts from £1,000 + VAT, and ongoing support plans start from £125 a month + VAT. To talk it through, Join the Circle.





