By the PrototypeVerdict Test Team
Our Base44, Bolt and Lovable links are referral links: if you sign up through one, Base44, Bolt or Lovable may reward us, at no extra cost to you. How we make money.
This is a pre-publish checklist for apps built with AI app builders (Lovable, Bolt, Base44 and similar). It is not a penetration test, a compliance certification, or legal advice. Platform scans help; they do not replace a review when you handle real client or personal data. To flag likely leaked keys or missing headers from pasted page source (nothing is uploaded), use our browser-only security check.
If you are still choosing a tool, start with vibe coding for beginners. If something already broke, see fix AI app builder errors and what actually broke in our tests.
Quick verdict
| Biggest risk we keep seeing | Database access rules (row-level security) that are missing, too open, or looping |
| Cheapest high-value check | Two signed-in accounts + one signed-out visit: each user must see only their own data |
| Do this before any real user | Run the builder’s security scan, fix Critical findings, confirm no secrets in frontend code |
| What a scan is not | A guarantee. Lovable’s docs say its scans “cannot guarantee complete security” and “do not replace a thorough security review.” Base44 says you are responsible for your app’s security and that the scan checks “common security problems.” Bolt’s docs describe what the audit covers and that some items may need your action; they do not say the scan misses findings. |
| Our lab note (Oct 29, 2026) | Lovable Deep scan on the portal: no findings (with the usual caveat). Base44 Check Security: 7 High dependency issues, 0 credits. Bolt project security audit: not run in our Free test |
Why AI-built apps leak data
These builders are fast because they generate a frontend, a database and access rules for you. Three structural facts drive most of the risk (Lovable states the same split in Security best practices):
- The browser is hostile. Anything in frontend code (including “hidden” admin buttons) can be inspected or bypassed.
- Your public database key is public on purpose. Protection has to come from row-level security (RLS) and server-side checks, not from hiding the key.
- Server/service-role code bypasses RLS. If a backend function uses privileged access, that code must check who is calling.
Our Oct 29, 2026 lab hit the middle of this stack on Bolt Free: after sign-up, login stuck with a 500 traced to recursive RLS policies on the profiles table. One fix message resolved that run (~84K tokens). We did not security-review Bolt’s elevated-privilege fix. Plain-English version: login and database guide — Step 4, full case WAB-01.
The checklist (print or copy)
Work top to bottom. Check a box only when you have evidence (screenshot, second account result, or scan report) — not when the AI said “done.”
A. Accounts and roles
- ☐ Named first admin in the original prompt (email or user id). We had to fix “Admins only / nobody is admin” in Lovable and set admin by hand in Bolt (add login and database).
- ☐ Sign-up path tested with a fresh account (email confirm, password rules).
- ☐ Sign-out ends the session (protected pages redirect after sign-out).
- ☐ Role changes (admin ↔ user) only possible for an admin, and only through a path you understand.
B. Row-level security (RLS) and data access
- ☐ RLS is on for every table that holds user or business data.
- ☐ Policies are not “any signed-in user can read/write every row.” Prefer owner/tenant checks.
- ☐ New tables added later also have RLS (easy to forget).
- ☐ Storage buckets (uploads) have matching rules — who can upload, list and download.
- ☐ You did not accept a fix that turns RLS off “to make it work.”
Two-account test (do this even if the scan is green):
- Create User A and User B with separate emails.
- As A, create a private record (intake, invoice, note).
- Sign in as B: you must not see A’s record or A’s files.
- Open the admin URL signed out: you should land on login or a public page, not data.
- Optional: open the app in a private window on another device after saving — we saw Lovable invoice data pass a same-preview reload but show $0.00 from another origin when storage looked browser-local (fix guide §5, WAB-04).
To see steps 1–4 run on one of our test apps, see our two-account access-control test of a Lovable app (one unpublished Lovable Pro client portal, fake data, checked through the app’s interface only).
C. Secrets and server-side logic
- ☐ No API keys, Stripe secret keys, service-role keys or passwords in chat history pasted into code,
.envcommitted to git, or frontend bundles. - ☐ Payments, emails and privileged API calls run in server-side functions, not in the browser.
- ☐ Prices and “is this user paid?” checks cannot be edited from DevTools on the client.
- ☐ For Stripe: follow the builder’s Connect form; never paste a live secret key into the chat as a normal message (Lovable warns about this in Connect your own Stripe account).
D. Builder security scans
| Tool | What to run | Plan notes (docs) | What we saw Oct 29, 2026 |
|---|---|---|---|
| Lovable | Publish dialog runs Quick scan; also run Deep scan from Security view or chat | Scans themselves are free; some fixes use the 10 free fix pool / credits (Security overview) | Deep scan on B1 portal: no issues flagged, with Lovable’s caveat that scans can’t catch everything. Dependencies panel showed 12 known issues across 53 packages |
| Bolt | Publish → Run security audit (preferred; docs say this path uses 0 tokens). Also Database → Security for RLS-focused checks | Full project audit: paid plans. Database Security tab: all plans (Check your project’s security, Database security settings) | Not run in our Free test (no audit button found / Free limitation). Recursive RLS bug was found by breaking login, not by an audit |
| Base44 | Dashboard Check Security | Free in our test (0 credits) | Ran ~1 minute, 5 checks, flagged 7 High dependency findings (zod + vite) and offered Fix all (we did not click it) |
- ☐ Critical / Error findings fixed or consciously accepted with a written reason.
- ☐ Re-run the scan after the last feature change before publish.
- ☐ Dependency High/Critical issues reviewed (Base44 and Lovable both surfaced package findings in our lab).
E. Publish settings and visibility
- ☐ App visibility is intentional (Private / Workspace / Public). In our Base44 test, new apps defaulted to Public visibility even though nothing was published — check before you share a link (publish guide).
- ☐ Preview URL and production URL both tested.
- ☐ Custom domain HTTPS works; old preview links not left world-writable with real data.
- ☐ For internal tools: require sign-in even if the link is “secret.”
F. After publish (first 24 hours)
- ☐ Create one more fresh user and repeat the two-account test on the live URL.
- ☐ Watch error logs / builder logs for permission denials and 500s at login.
- ☐ Confirm backups or export path (export and backup guide).
- ☐ If you handle accountant/client PII: keep prototypes on fake data; talk to your compliance advisor before real SSNs or tax data — see vibe coding for accountants.
Tool-by-tool: where the security controls live
Lovable
- Review policies: More → Database → Manage → Policies (Control who can access your app’s data).
- Scans: project Security view; Quick scan on publish; Deep scan on demand (Security overview).
- Pre-publish list from Lovable docs: no secrets in frontend; validation server-side; RLS configured and tested; auth enforced server-side; Critical findings addressed (Security best practices).
- Optional: workspace Security center (Business/Enterprise), Wiz, Aikido pentest connectors (third-party cost).
Bolt
- Full audit: Publish → Run security audit (up to 30/day, resets midnight UTC; prompting an audit in chat uses tokens, the Publish button path does not — per Bolt docs).
- Database-only: database icon → Security (Supabase-advisor-style warnings such as RLS disabled or “RLS Policy Always True”).
- If you use your own Supabase, also check Supabase Advisors — Bolt’s DB tab uses the same class of checks (Database security settings).
Base44
- Run Check Security from the app dashboard (our Oct 29 run: dependency-focused findings, Fix all offered, 0 credits).
- Review App Users, roles and visibility before publish.
- Stripe and other paid integrations: confirm plan requirements (Stripe docs: Builder+) and that test mode is used until go-live (Setting up Stripe).
Messages that help (not magic spells)
Use after you have a failing two-account test or a scan finding. Adapt; we have not A/B tested this wording.
RLS too open:
Users can read each other’s records on [table]. Turn on row-level security if it is off. Policies must allow each signed-in user to read and write only their own rows (match auth user id to the owner column). Do not disable RLS. Change nothing else.
Admin lockout:
The /admin page says Admins only and I have no way to become an admin. Make [admin email] an admin and show the Admin link only to admins. Do not weaken security rules. Change nothing else.
Recursive policy / login 500:
After sign-up, login gets stuck with a 500. Check whether security policies on the profiles table refer back to themselves. Fix without turning RLS off, keep each user limited to their own data, explain the change, and change nothing else.
More fix patterns: fix AI app builder errors.
What we are not claiming
- We are not saying Lovable, Bolt or Base44 are “secure” or “insecure” as platforms.
- We did not run Bolt’s paid security audit, external pentests, or production load tests.
- Dependency findings (Lovable 12 issues; Base44 7 High) were reported by the vendors’ scanners in our unpublished test apps; we did not verify exploitability.
- Outside research (for example press coverage of vibe-coded app vulnerabilities) is out of scope for this checklist.
For publish mechanics (domains, badges, visibility), use how to publish an AI-built app. For rescue after a bad change, use the app builder troubleshooter.
FAQ
Is Lovable safe for client data?
Lovable provides built-in Quick and Deep scans, but you own row-level security, secrets and what you put in the app (Security overview). For real client data, run Deep scan, pass the two-account test, and get a qualified review if the data is sensitive. We are not your compliance advisor.
What is row-level security (RLS)?
A database rule that decides which rows a request may read or change. With the public anon key anyone can try to query your tables; RLS is what blocks cross-user access (Lovable database access; Bolt database security).
Do Lovable / Bolt / Base44 security scans cost credits or tokens?
Lovable: running Quick/Deep scans is free; some automated fixes use the free-fix pool or credits (Security overview). Bolt: Publish → Run security audit uses no tokens on paid plans; asking in chat does use tokens (Bolt security). Base44 Check Security used 0 credits in our Oct 29 test.
Why did our Bolt login return a 500?
In our Free test, recursive RLS policies on profiles caused a login 500; one fix used ~84K tokens. That is one cause among many. See what actually broke — WAB-01 and fix guide.
Can I publish if the scan still shows warnings?
Vendors differ. Lovable workspaces can optionally block publish on Critical findings. Bolt and Base44 still expect you to read findings. Our rule: fix Critical/Error access-control issues before real users; document any Warning you accept.
Is a “secret” preview link enough protection?
No. Treat unauthenticated preview URLs as public. Require sign-in for private data and set visibility deliberately (publish guide).
Should accountants put client files in a vibe-coded portal?
Only with fake data until access control is proven, and only after your firm’s own security/compliance review. Start with vibe coding for accountants and build a client portal with AI.
Related pages
- Add login and a database to an AI app
- What actually broke in our tests
- Fix AI app builder errors
- Publish an AI-built app
- Export and back up AI app code
- App builder troubleshooter
- How we test
Update history
| Date | Change |
|---|---|
| Nov 9, 2026 | Vendor security docs checked (Lovable, Bolt, Base44); lab observations from our Oct 29, 2026 tests. |
Vendor security docs checked: Nov 9, 2026. Lab observations: Oct 29, 2026 (Lovable Pro, Bolt Free, Base44 Builder, Softr Basic). Published: Nov 9, 2026. Updated: Nov 16, 2026.