Security of AI-built apps: what is visible from the outside
Libraries with known vulnerabilities and five missing protective layers, measured across 3,728 apps with a DACH connection. With count, base and a clear limit: what the outside view cannot see.
6% of the 3,728 scanned AI-built apps shipped a front-end library with a publicly known vulnerability (6% (240 of 3,728)). 98% had no Content-Security-Policy, 43% no security.txt. Source: decivo AI App Radar DACH, run “2026-Q4-recall”. The check ran from the outside only; database and login were not part of the measurement.
Key points
- 6% of the 3,728 discoverable AI-built apps with a DACH connection (240 apps) shipped a front-end library in a version with a publicly known vulnerability.
- 98% of the 3,728 discoverable AI-built apps with a DACH connection (3,653 apps) had no Content-Security-Policy.
- 43% of the 3,728 discoverable AI-built apps with a DACH connection (1,585 apps) had no security.txt, that is, no named contact for vulnerability reports.
- Missing protective layers are frequent, but no measure of severity. They are missing on many well-kept professional sites too.
The figures
The first row is a defect of severity “high”. The other five are missing protective layers: one header or one DNS record each.
- Shipped library with a known vulnerability
- No Content-Security-Policy
- No clickjacking protection
- DMARC in monitoring mode only
- No security.txt
- No CAA record
Sentences to quote
You can use these sentences word for word. Each one names share, base and count. Please include the source line below.
6% of the 3,728 discoverable AI-built apps with a DACH connection (240 apps) shipped a front-end library in a version with a publicly known vulnerability.
98% of the 3,728 discoverable AI-built apps with a DACH connection (3,653 apps) had no Content-Security-Policy.
90% of the 3,728 discoverable AI-built apps with a DACH connection (3,342 apps) had no protection against being embedded in other sites (clickjacking).
43% of the 3,728 discoverable AI-built apps with a DACH connection (1,610 apps) had DMARC set to “p=none”, so forged mail is still delivered.
43% of the 3,728 discoverable AI-built apps with a DACH connection (1,585 apps) had no security.txt, that is, no named contact for vulnerability reports.
99% of the 3,728 discoverable AI-built apps with a DACH connection (3,697 apps) did not restrict which authorities may issue certificates for their domain (no CAA record).
What is actually urgent
The library with a known vulnerability is the finding that counts. Whatever an app ships to every visitor can be read by anyone, including the version of a bundled library. If a public advisory exists for it, the flaw is documented. The rule only fires for a reliably detected version, matched against the open database OSV.
The Content-Security-Policy tells the browser which sources it may run scripts from. It is a second line of defence, not proof of a flaw. That it is missing in almost all apps mainly shows that no builder sets it by default.
A security.txt as defined in RFC 9116 names the address someone can report a discovered flaw to. It costs two lines of text. DMARC in monitoring mode and a missing CAA record are hygiene: quick to set, rarely acute.
What the figures do not say
The radar measures what can be found. The 3,728 apps are not a random sample of all AI-built apps. The correct phrasing is “x% of the 3,728 discoverable AI-built apps with a DACH connection”, never “x% of all Lovable apps”.
The check runs from the outside and without logging in: what the page ships and what a browser observes while loading. The radar does not see database rules, roles or anything behind the login. A share says for how many apps a rule fired. It does not say whether the finding holds in each case, and it is not legal advice.
The report deliberately does not cite everything. At the time, for fonts, trackers, contacted services and findings in shipped code a mere mention also counted, for example in an embedded third-party script. Since methodology 2026.43 only what the page demonstrably loads or contains in its own code counts; exception: if the page ships no code of its own, the rules for database write calls and role comparisons check an embedded third-party script. These shares may be lower when measured again. Those shares are in the dataset, but not in this text.
The figures are decivo's own survey. decivo itself offers to fix such findings. They have not been independently verified so far; the dataset is open for that purpose.
The biggest risks of AI-built apps often sit in the database: tables readable without logging in. The radar does not measure exactly that, because it queries no database. These figures therefore describe the visible surface, not the whole risk.
The basis is the run “2026-Q4-recall”, closed on 8 November 2026. Across all 14 runs since September 2026 the radar has scanned 4,402 apps; the report uses the single run because it has one methodology and a fixed address.
Check your own app
Two things can be done within an hour: raise the reported library to the patched version and place a security.txt under /.well-known/. A Content-Security-Policy is best introduced in report-only mode first and tightened afterwards.
The free scan names the affected library and its version. Checking the database as well needs a scan with access. All figures in context are in the AI App Report DACH 2026.
Sources
Every share comes from the radar run named below. Its fixed address shows exactly these figures; the legal sources are the statutes themselves.
- decivo: AI App Radar DACH, run “2026-Q4-recall”, 3,728 apps, methodology 2026.37 (CC BY 4.0)
- decivo: raw figures of the run as JSON (CC BY 4.0)
- decivo: AI App Report DACH 2026, November 2026 edition
- OSV: Open Source Vulnerabilities
- RFC 9116: A File Format to Aid in Security Vulnerability Disclosure (security.txt)
- RFC 8659: DNS Certification Authority Authorization (CAA)
- MDN: Content Security Policy (CSP)
Whether your app is part of these shares, the free scan shows in a few minutes: from the outside, without access, with evidence for each finding.
Check your app for freeQuestions about security
Are AI-built apps insecure?
That cannot be said in general from the outside. What can be measured: 6% of the 3,728 discoverable AI-built apps with a DACH connection (240 apps) shipped a front-end library in a version with a publicly known vulnerability.
How many AI-built apps have no Content-Security-Policy?
98% of the 3,728 discoverable AI-built apps with a DACH connection (3,653 apps) had no Content-Security-Policy.
Does the radar check the database?
No. The radar queries no database, creates no account and tests no login.
Related articles
- AI code quality
Accessibility of AI-built apps: radar figures 2026
The most frequent finding in the AI App Radar DACH: serious accessibility violations in the fully loaded page. Figure, method, limits and a sentence to quote.
- AI code quality
Privacy in AI-built apps: radar figures 2026
Trackers before consent, cookies without agreement, missing legal notice: what the AI App Radar DACH measured in AI-built apps, with count, base and sentences to quote.
- Vibe Code Rescue
Lovable app not working anymore? The honest emergency guide
Your Lovable app breaks on every click and the AI agent makes it worse? Common errors and their causes, what revert and “Try to fix” really do, what to check yourself in 5 minutes and when a rescue beats a rebuild.
Where does your app stand by comparison?
The decivo Scan checks your app from the outside, with the rules as they stand today, and names every finding with evidence.
30 minutes, free. You talk to us directly.