Lovable apps expose passport details and more than 200K private messages

An AI-powered no‑code platform called Lovable is under fire after research showed thousands of apps built on it exposed sensitive user data, including passports, emails and private messages, due to misconfigured Supabase row‑level security and client‑side database access. Commenters argue this is less a novel AI failure and more an acceleration of long‑standing industry problems: insecure defaults, “move fast” culture, and pushing complex security responsibilities onto non‑technical users while marketing that “no coding” is required. Many call for stricter liability, safer-by-default architectures, and platforms that simply refuse to deploy apps with critical privacy flaws, especially as AI tools both generate vulnerable code and make exploitation easier.

Scope of the Lovable / Supabase Exposure

  • Analysis of 159k Lovable projects, ~34k using Supabase; ~24k loaded DB records on page load without auth.
  • Email addresses found in >10k tables across 7.4k sites; an LLM-based classifier judged nearly 4k sites as leaking privacy‑sensitive data (16% of Lovable+Supabase sites).
  • Researcher stresses vulnerabilities were trivial to exploit (basic network inspection, simple extra queries), not advanced hacking.

Responsibility: Platform vs. Database vs. App Owners

  • Many argue Lovable bears major responsibility: it markets “no technical knowledge needed,” yet disclaims security liability and exposes non‑technical users to complex Supabase/RLS pitfalls.
  • Others note app owners are formally the controllers under GDPR and made poor choices (e.g., collecting passports on such a platform).
  • Supabase is criticized for historically insecure defaults (frontend anon keys with broad access, RLS complexity, no rate limiting) that are easy to misconfigure, especially for “vibe coders.”

AI, “Vibe Coding,” and Software Quality

  • Several see this as emblematic of AI‑assisted “vibe coding”: polished apps by people who don’t understand auth/authorization, vastly enlarging the attack surface.
  • Others emphasize these are old cultural/incentive problems (“move fast,” growth-first) that AI has simply poured gasoline on.
  • Some defend AI as a powerful tool if used by experienced engineers with good governance; the problem is treating AI as a replacement for that expertise.

Security Culture, Liability, and Regulation

  • Multiple comments call for strong liability and GDPR enforcement: stiff fines, easier user compensation, possibly “corporate death penalties” for repeat offenders.
  • Others are pessimistic: users tolerate abuse, companies optimize for KPIs, and only catastrophic incidents may trigger real regulation.

Researcher Ethics and Legality

  • Big debate over whether sampling thousands of tables crosses from “proof of concept” into legally risky “mass exfiltration.”
  • One side: only minimal pages were fetched, data deleted, lawyer consulted; necessary to quantify blast radius.
  • Opposing side: once you download more than needed to show a hole, you’ve created an additional breach and legal exposure under regimes like GDPR; independent researchers should stop earlier or work under contract/NDAs.
  • No consensus; lawfulness and best practice remain contested and jurisdiction‑dependent.