Insecure by Default: Our Research on AI-Generated Web Applications

AI coding tools have democratized software development. Non-technical founders can now ship production applications in hours. But at what cost to security?

We conducted the largest empirical security study of AI-generated web applications to date, analyzing 603 production apps built with Lovable, Bolt, and Replit.

Key Findings

Security Header Analysis

Security headers are the first line of defense for web applications. They instruct browsers on how to handle your content securely. AI coding tools almost universally fail to configure them.

Other Findings

Source Map Exposure

34.1% of applications had source maps exposed in production. Source maps are debugging files that reveal your entire original source code—including comments, variable names, and application logic—to anyone who views them.

Client-Side Secrets

We detected API keys and sensitive configuration exposed in client-side JavaScript bundles. Common patterns included:

• Supabase anon keys (expected, but often with overly permissive RLS)
• Firebase configuration objects
• Stripe publishable keys (legitimate) alongside secret keys (not legitimate)
• Third-party API keys for services like SendGrid, Twilio, OpenAI

Backend-as-a-Service Misconfiguration

Applications using Supabase and Firebase frequently had:

• Row Level Security (RLS) disabled or overly permissive
• Anonymous authentication enabled when not needed
• Storage buckets with public read access
• Realtime subscriptions without proper authorization

Platform Comparison

The consistency across platforms is notable: the security gaps are systematic, not platform-specific. This suggests the issue lies in how AI models are trained and what they optimize for (functionality over security).

Implications

For Developers and Founders

If you've built with AI coding tools, assume your application needs a security review. The vulnerabilities are predictable and fixable, but only if you know to look for them. A quick scan before launch can prevent embarrassing (or costly) security incidents.

For AI Coding Tool Providers

There's an opportunity to differentiate by implementing secure defaults:

• Auto-inject security headers in deployment configurations
• Warn when secrets appear in client-side code
• Generate stricter BaaS security rules by default
• Disable source maps in production builds automatically

For Security Professionals

Vibe-coded applications present a predictable attack surface. When assessing AI-built apps, prioritize:

1. Security header injection and configuration
2. Client-side JavaScript analysis for exposed secrets
3. BaaS permission and RLS policy review
4. Source map and debugging artifact exposure

Methodology

Discovery

We identified production applications through:

• Platform galleries (launched.lovable.dev, bolt.new/gallery, madewithbolt.com)
• Certificate Transparency log mining for platform-associated subdomains
• Fingerprinting to confirm AI-generated origin

Scanning

Each application was analyzed for:

• HTTP security headers (presence and configuration)
• Client-side JavaScript bundle content
• BaaS configuration indicators
• Technology fingerprinting and exposure

Scoring

Applications received a weighted score (0-100) based on the presence and proper configuration of security controls, mapped to letter grades (A-F).

Citation

@article{winters2026insecure,
  title={Insecure by Default: A Cross-Platform Security
         Analysis of AI-Generated Web Applications},
  author={Winters, Kirby and Merrill, Todd},
  year={2026},
  publisher={CyberSavi}
}