Septr Research

What AI-generated apps actually ship.

We study source code and public deployments with one rules engine, publish the raw methodology, and separate harmless signals from credentials that can access data or spend money.

Explore the repository See Septr in action

The live deployment study

The dangerous part is what reaches production.

We probed publicly deployed applications across Vercel, Netlify, Cloudflare Pages, Lovable, and Replit. The study found real credentials in client-side JavaScript, where every visitor can retrieve them.

1,248unique public deployments probed

11critical API keys in client bundles

4Supabase service-role keys exposed

7OpenAI keys exposed in production

The distinction matters. A credential in a repository is a source-code finding. The same credential compiled into a live bundle is a production exposure served to the internet.

Published charts

Evidence you can inspect.

These charts are generated from the published, anonymized datasets. Read the full study for sampling, adjudication, confidence intervals, and limitations.

Committed environment files across the Lovable corpus — Committed environment files in the Lovable corpus.

Blast radius tiers for findings — Blast radius: data access, cost, bounded exposure, and advisory findings.

Comparison across AI-generated app populations — Four populations, one rules engine, one adjudication standard.

Methodology

Precise enough to be useful.

One engine

The same Septr rules engine scans source archives and live responses, so the two studies can be compared without changing the detector.

Human verification

Findings were reviewed with surrounding code. Placeholder values and public-by-design keys are separated from real exposures.

Ethical publication

We publish aggregates and anonymized rows only. No repo-to-credential mapping and no confirmed-live credential testing.

Read the live study Read the corpus study