// Incident landscape
Security incidents in AI-built SaaS in 2026
Quick answer
If you thought "a small AI-built app is not a target", 2026 proved otherwise, with numbers. This is not security-vendor FUD: these are public incidents and at-scale studies that show a single pattern. The product ships fast with AI, the flows that sustain revenue (login, billing, customer data) go without a human review, and the leak shows up days after launch, not years.
This guide gathers the real 2026 cases and studies in one place, with sources, so you calibrate risk by evidence, not fear. The practical takeaway: reviewing the sensitive flows before (or right after) launch is the cheapest step there is, and that is exactly where RET fits.
The 2026 numbers
Each row below is a public case or an at-scale study from 2026. The recurring theme: sensitive data exposed by loose access control, early, in AI-built products.
| Case / study | What happened | Scale |
|---|---|---|
| Moltbook (Jan 2026) | A 100%-AI-built social network exposed API tokens and e-mails | 1.5 million tokens + 35k e-mails within 72h of launch |
| Lovable postmortem (Feb-Apr 2026) | Public projects’ source code and chat history readable by any logged-in user | 11 weeks of exposure |
| RedAccess study (May 2026) | A scan of vibe-coded apps (Lovable, Base44, Replit, Netlify) found leaks | ~5,000 of 380,000 apps; 40% with medical/financial/corporate data |
| GuardMint (Q1 2026) | Vibe-coded apps audited with at least one hallucination/missing-context vuln | 91.5% (183 of 200) |
| OX Security (2026) | AI-generated code shipping with a vulnerability | 62% of code |
| Supabase CVE-2025-48757 | RLS off exposed tables via the anon_key | 303 endpoints across 170 apps |
The pattern behind the numbers
None of these incidents needed an exotic exploit. The common denominator is access control: RLS off, an unauthenticated endpoint, a public id becoming a key, a permission only in the front-end. AI generates the happy path fast and leaves the access boundary loose, because access does not show up in the demo, it shows up in the incident.
- Loose access (RLS off, unauthenticated endpoint, IDOR) is the dominant cause.
- Time to leak is short: days or weeks after launch.
- Sensitive data (payment, health, PII) is usually in play from early on.
What to do about it
The 2026 lesson is not "stop using AI", it is "review the flows that sustain revenue before they become a headline". Start with the Promptbook (R$147) to catch the obvious yourself, move to Risk Review when a signal appears in login, billing or data, and turn on AuditMySaaS so you are not blind between deploys. The cost of reviewing is a fraction of the cost of becoming a case on this list.
Want to know whether this happens in your system?
Manual pentest for web applications, APIs and apps, from R$9,500, with the Sprint run within 7 business days. We refund 50% if we find no flaw.
Frequently asked questions
My app is small, am I really a target?
The 2026 cases show size does not protect: Moltbook leaked 1.5 million tokens within 72h of life, and the RedAccess study found leaks in thousands of small apps. Automated scanners sweep new apps at scale; the target is the exposed data, not fame.
Is this the AI tool’s problem or mine?
Both. The tool accelerates and leaves access control loose by default, but the app and the data are yours, so the correct policy is the publisher’s responsibility. The good news is most of these flaws are quick to fix once you know where to look.
How do I know if I am exposed?
By testing the access flows: what the public key reaches with no login, whether a swapped id returns another customer’s data, whether an endpoint responds with no auth. The Promptbook guides that check and Risk Review does the human read when the signal is serious.
Who runs the review at RET?
Gabriel Lima Ferreira, founder and lead pentester, with an authorized scope and reproducible evidence. RET has over 100 real vulnerabilities reported with responsible disclosure.
Keep reading
Threat and defense
MCP tool poisoning: how to prevent it
Practical guide
Multi-tenant isolation in SaaS
Threat and defense
Credential stuffing and MFA on the admin panel
Comparison
Reviewing your AI SaaS yourself vs human review
By tool
Security for apps built with Lovable
By city
SaaS security and pentest in Recife (PE)