Orca Security: What Is AI AppGen Security?

What it is

Orca's definition piece on "AI AppGen security" — governing running applications that AI app builders (Replit, Supabase, Vercel, Lovable, Base44, Netlify) generate and deploy outside the development pipeline. Beyond the category argument, it carries concrete configuration mechanics and a zero-cost discovery method the summary omits.

How it works — the RLS trap the summary misses

The summary notes SQL-editor-created Supabase tables don't auto-enable row level security. The page adds the second trap that follows: once RLS is on, no data reaches the API through a publishable key until policies exist, so a builder hitting an empty screen asks the assistant to fix it — and a policy granting select to the anon role with a condition of true makes the table readable by anyone again. Views compound this: Postgres creates them with security definer, and only Postgres 15+ lets the creator restore RLS enforcement with security_invoker = true. Supabase documents the publishable key as safe to expose in client source; secret/legacy service-role keys bypass RLS entirely and must never reach clients.

Operational detail on Replit: changing an app's access level "requires unpublishing your app first" — the fix takes the app offline. Vercel's Advanced Deployment Protection (private production deployments) is Enterprise-included, or +$150/month on Pro.

What to do

• Discover the way the researchers did, first. RedAccess located thousands of exposed apps with ordinary search queries against the vendors' own hosting domains (apps hosted on the vendor domain rather than a custom one). Costs nothing; gives a starting inventory before any tooling decision.
• Per platform, record account type, plan, protection scope, and what it excludes — not "is security on." On Replit, which workspace published it; on Vercel, which plan pays for production protection.
• Move enforcement into the database. RLS on generated tables survives the next prompt; assistant-written app logic can be rewritten by the following prompt. Check each policy names a role and restricts rather than permits.
• Provide an org workspace — it flips the Replit publish default from Public to private without relying on anyone remembering.
• Attach an owner and review date to every generated app. Report inventory and ownership monthly, data/identity reach continuously.

Caveats worth noting

The 5,000+ figure counts only apps on vendor domains; Zvi estimates thousands more sit on builder-purchased domains, with ~2,000 of those found appearing to reveal private data on closer inspection. The vendors' on-record responses (Replit, Lovable, Base44/Wix) all frame exposure as a user configuration choice, not a platform vulnerability — i.e. shared-responsibility by design, not a bug anyone will patch for you.