I built a free tool that attacks a Supabase app as one customer to see if it can read, change or delete another customer's data (built with Claude Code)
retained text
I built this, so weigh my enthusiasm accordingly. It's free and MIT licensed. I'm not a programmer: I'm a product owner in South Africa who builds by directing Claude Code. **Why** I run my own rental-management app on Supabase. The question that worried me: can one of my customers see or change another customer's data? Row level security is easy to get wrong, and I can't read the policies well enough to be sure. So I had Claude Code build something that checks by trying, not by reading. **What it does** `tenant-leak-check` rebuilds your app's database on your own machine from your migration files (Docker, local only; it refuses anything that isn't a local address and needs no keys). It creates two made-up customers with made-up data. Then it really attacks, as customer A and as a visitor who isn't signed in: - tables: read, add, change, move a row to the other customer, delete - storage: list, download, upload, overwrite, delete the other customer's files - database functions: called with customer B's ids You get an HTML, Markdown and JSON report with the evidence (what was sent, what came back) and a plain-English fix for each leak, plus exit codes so CI can fail the build. Migration mode runs the app before and after a new migration and flags new leaks and risky changes: policy removed, RLS switched off, column dropped, with line numbers. There's also a Claude Code skill: you say "run the leak check" and it runs it and explains the report. **What I checked it on** - A demo app with 19 planted leaks: it finds all 19. That's 48 failed checks, none missed, none extra, and the tests require an exact match. The fixed version of the same app comes out clean. - My own rental-management app: 669 cross-customer checks, 0 leaks, plus 3 smaller findings worth fixing. - A deliberately bad migration I made up for that app: 6 new leaks and 4 flagged lines. **How Claude Code was used** Claude Code (Opus 5.5) did all the building, in four sessions. I wrote the brief and the rules: local only, never weaken a test, and a check can never "pass" unless it was really carried out; otherwise it must say "not tested" or "unclear" and give the reason. The planted-leak demo was the key part, because it showed the tool catching what it claims to catch. **Limits** - It's an automated isolation test, not a security certification, audit or penetration test. "0 leaks" means none found in what was tested. - It tests a local copy built from migration files, not your live database. Anything changed by hand in production is invisible to it. - It doesn't test your app code, Edge Functions, anything using the service-role key, Realtime or webhooks. - Two made-up customers with a few rows each, so rules that depend on lots of data may not get exercised. - Database functions are called once with automatic arguments. If one refuses because the input was invalid, the report says so, and what it does with a realistic request isn't proven. - Needs Node 24, Docker and the Supabase CLI. Early version (0.1.0), only tried on the demos and my own app. **Try it** https://github.com/seanjohnson-dev/tenant-leak-check The README walks through it step by step, and the demo apps run in a few minutes. If you'd rather have me run it on your project and explain the report, message me. More about what I do: https://swjdigital.com **A question** If you run a Supabase app with more than one customer: which leak shapes should it try that it doesn't? And if it gets something wrong on your project, tell me.
document fetched 2026-11-03 00:41 · canonical https://reddit.com/r/ClaudeCode/comments/1xlt2n7/i_built_a_free_tool_that_attacks_a_supabase_app