Blog

Security research, case studies, and guides for AI-built apps.

SupabaseOctober 20266 min read

Supabase's October 30, 2026 Data API change: what breaks in Lovable, Bolt and Cursor apps

From October 30, 2026 new tables in existing Supabase projects are no longer exposed to the Data API automatically. Existing tables keep working; new tables need explicit GRANTs plus RLS. Here is what changes, what breaks, and the exact SQL.

SupabaseOctober 20266 min read

16,326 Supabase databases were publicly readable: what UpGuard found and how to check yours

In September 2026 UpGuard reported 16,326 Supabase databases with publicly readable tables. The cause was not a Supabase breach but tables without Row Level Security. A five-minute self-check, with SQL.

LovableOctober 20265 min read

Is Lovable's built-in security scan enough? What Quick Scan and Deep Scan check — and what they don't

Lovable runs a Quick Scan automatically when you publish and offers an on-demand Deep Scan. Both are static checks of your project. Here is exactly what they cover according to Lovable’s docs, and the gaps you still need to close.

CVEOctober 20265 min read

CVE-2025-48757 explained: the Lovable Row Level Security vulnerability and what it teaches

CVE-2025-48757 (CVSS 9.3) described generated Lovable sites whose database policies let unauthenticated attackers read or write tables. What the CVE says, why it is disputed, and the checks that prevent the same class of bug.

SupabaseOctober 20265 min read

Supabase is retiring anon and service_role keys by the end of 2026: how to migrate to publishable and secret keys safely

Supabase is deprecating the legacy JWT anon and service_role keys by the end of 2026 in favor of sb_publishable_ and sb_secret_ keys. What each key can do, where each may live, and a migration that does not leak the secret key.

SecretsOctober 20265 min read

Why your OpenAI or Anthropic API key ends up in your frontend bundle — and how to find it first

AI-built apps often call LLM APIs straight from the browser, which puts the provider key in JavaScript anyone can download. How the leak happens with VITE_ and NEXT_PUBLIC_ variables, how to find it, and the server-side pattern that fixes it.

SupabaseOctober 20265 min read

Protecting Supabase Edge Functions that call LLMs: verify_jwt, rate limits and webhook signatures

An Edge Function that calls an LLM is a public, billable endpoint. Supabase verifies JWTs by default — unless verify_jwt is turned off. How to keep auth on, rate-limit per user, and handle the functions that must be public.

MCPOctober 20267 min read

MCP server security in 2026: tool poisoning, rug pulls, token passthrough — a practical checklist

Model Context Protocol servers give AI agents tools — and attackers a new injection path. What tool poisoning and rug pulls are, what the 2026-07-28 MCP specification requires, and a checklist for using and building MCP servers safely.

PaymentsOctober 20265 min read

Stripe webhook signature verification in AI-generated apps: the check that stops free orders

An unsigned Stripe webhook lets anyone mark an order as paid. How Stripe signs events, why AI-generated handlers break verification, and a correct handler with replay protection.

ResearchOctober 20267 min read

AI-built app security report 2026: what 799 live apps show (Sekrd dataset)

Sekrd’s anonymized dataset of 799 publicly reachable hosts scanned March–September 2026, including 174 on lovable.app. 4.4% had a critical finding, 81.5% served no Content-Security-Policy, 60.3% had no privacy-policy link. Method, numbers and limits.

ResearchApril 20268 min read

100 vibe apps, one scanner: what a random sweep of the AI-built web actually finds

AI coding tools ship faster than any IDE in history. Platform defaults haven't caught up. We scanned 100 random production apps from 12 AI-coding hosting platforms: 5% had critical live leaks (Stripe keys in the browser, full Supabase table exposure), 57% shipped with no Content-Security-Policy at all. The fix is continuous scanning in CI — the same pre-commit discipline that stops typos from reaching production.

ResearchApril 20267 min read

98 vibe-coded apps, 967 findings: the security posture is the platform, not the app

We ran ~100 scans across vibe-coded apps and indie deployments from two different cohorts. 96% shipped, 4% got BLOCK, 21% bounced off their own WAF. Inside each cohort, scores clustered so tight it surprised us. The real risk doesn't come from the template — it comes from the 3–6% of apps where the developer reached past the template.

ResearchApril 20268 min read

50 sites, 444 findings: the 2026 state of shipping-fast security

April 2026. We scanned 50 production URLs — vibe-coded AI tools, small-business WordPress, auth-SaaS homepages. 41 scored, 9 bounced off modern WAF and we respected it. 88% SHIP, 12% BLOCK, median 76, 444 findings. Every site had findings. Here's what the shape of those findings says about how the web is being built right now, and why 'ship fast' can't keep ignoring the security layer.

Deep ScanApril 20265 min read

5 security checks a URL-only scanner will never catch

A URL scan checks your front door. Deep scans check whether the windows are locked. Here are the five classes of findings that only show up when you plug in your Supabase, Firebase, or MCP credentials.

SupabaseApril 20264 min read

Your Supabase RLS says "enabled." Here's why your database is still open.

A Row Level Security policy with USING(true) passes every dashboard check and leaks every row. We've seen it in production apps, in our own test projects, and in the Supabase docs themselves. Here's what it looks like and how to find it before someone else does.

FirebaseApril 20264 min read

"allow read, write: if true" — the Firestore rules gap nobody fixes in time

Firebase test-mode rules give you 30 days. A lot of apps get a year. We've scanned projects with wide-open Firestore, wide-open Storage, and surprisingly tight rules that just needed a matching deny-all fallback. Here's what the audit actually checks.

ResearchApril 20268 min read

We scanned 49 of the internet's biggest sites. Here's what we found.

GitHub, Google, Slack, Stripe, Notion, Vercel, and 43 others. 89% passed. 5 got BLOCK. Every single site had findings. The average score was 80/100. Full data and methodology inside.