2025-11-12

What Escape.tech published about 5,600 vibe-coded apps

In October 2025 Escape published The State of Security of Vibe Coded Apps and a methodology post by Nohé Hinniger-Foray, Gwendal Mognier, and Alexandra Charikova (29 October 2025). The PDF is hosted here. The numbers below are theirs. LeakSweep has not repeated the study.

The methodology post says the team analysed over 5,600 publicly available applications and identified more than 2,000 vulnerabilities, 400+ exposed secrets, and 175 instances of personal data, including medical records, IBANs, phone numbers, and emails. After discovery they describe 14,600 assets: 5,600 web applications, 1,280 API services, 6,500 hosts, and 1,103 schemas. They say scans ran in a passive mode, without destructive operations, and that the results are a lower bound.

The PDF's executive summary says the same assessment covered over 5,600 publicly available applications, that nearly 60% contained critical security flaws, and that 34,232 vulnerabilities were found, including nearly 2,000 high-impact issues. It lists 400+ exposed secrets, including 6 GitHub tokens, 7 OpenAI API keys, and 2 admin payment API keys, plus 175 instances of personal data. It names misconfigured permissions, particularly Supabase row-level security, as the primary cause, rather than classic injection bugs. Platforms named include Lovable, Base44, Vibe Studio, Create.xyz, and Bolt.

The findings section of the same PDF also says "1.4K apps tested after cleaning" and separates ">400 instances of leaked secrets" from "400+ exposed public state-altering endpoints". Those are not the same statistic. The leaked-secret breakdown on that page lists 2 admin payment API tokens, 6 GitHub token leaks, 33 plaintext passwords, and 7 OpenAI API keys. Read the PDF rather than a paraphrase that collapses these figures.

Two points in the findings section matter for anyone shipping a generated app. First, Escape says the Supabase anon token is expected in the browser and is not, by itself, a vulnerability. Second, they say large language models often get database permissions wrong, and they connect one GitHub token leak to insufficient row-level security described by CVE-2025-48757. The report also cites research from the Center for Security and Emerging Technology at Georgetown on bugs in model-written code. That citation is Escape's; go to the PDF if you want their wording.

LeakSweep is built around the part of this that an owner can check without hiring a reviewer: prove the domain is yours, look at public files and keys, and test whether the public database key can count or write rows. It will not reproduce Escape's study, and a clear LeakSweep report is not a claim that your app would have passed theirs.