Adrian Cisneros Contact

I scanned 2,487 AI-built apps

October 2026

A security scan of 2,487 public AI-built apps found databases anyone could write to.

The headline I scanned 2,487 AI-built apps above a code block with a policy set to true

I scanned 505,071 files from 2,487 public GitHub repos built with AI coding tools.

Row level security was turned on for 11,682 of 11,952 tables. Even so, 109 of 442 Supabase apps, about a quarter, still had a table anyone could write to, often through a rule set to true.

Only 38 repos, 1.5 percent, had a live server-side secret committed. 156, or 6.3 percent, had at least one serious finding.

I checked precision by hand on samples. I tested no credentials and named no repos. I also wrote a one-file checker people can run on their own repo.

  • Scanned505,071 files in 2,487 repos
  • WhenOctober 2026
A grid of 442 squares, one per Supabase app, with 109 colored amber
109 of 442 Supabase apps had a table anyone could write to.

More in Research

Paper 46, explainedAn interactive explainer of my gesture radar paper, for people who have not read it.A line chart of detection score for six detectors as up to 20 percent of stored numbers are set to zeroA fly-brain detector for factory control attacksA pilot test of a brain-inspired attack detector for factory controllers on failing hardware.
A flow diagram of one research day in the agent town, where sealed held-out data feeds only the replicatorA town of AI agents looking for brain findingsA survey and a prototype town of AI agents that test brain-data claims on sealed data.Teaching a gesture radar a new user without backpropagationOctober 2026A short paper for IEEE-EPIC 2026 on retraining a radar gesture model with forward passes only.
More ResearchAll projects

Want something like this?

Answer a few quick questions and see an estimate in about a minute.

Get an estimate

or email contact@adriancisneros.dev