LUCID

AI code compiles. It doesn't work.

LUCID finds the bugs your AI assistant hallucinated into existence. Adversarial verification for every implicit claim in your code.

Bolt.new
0/100
Lovable (App 1)
0/100
Lovable (App 2)
0/100
Replit
0/100

We adversarially verified code from 4 platforms. 21 bugs found. 0 passed.

The worst bugs we found

Highest-severity finding per platform. Selected from 21 total.

Bolt.new42/100
critical

Broken config bootstrap — app non-functional out of the box

ensureRuntimeSupabaseConfig is never called before React renders. All Supabase API calls fail with undefined URL/keys.

Lovable (App 1)42/100
critical

Admin routes with zero authentication guards

All /admin/* routes (users, monitoring, configuration, tools) are defined with no auth checks. Any user can navigate directly to admin panels.

Lovable (App 2)42/100
critical

iframe sandbox disabled — arbitrary script execution

BrowserPreview.tsx has sandbox attribute commented out. No restrictions on embedded content, allowing arbitrary script execution.

Replit32/100
critical

Scene analysis calls non-existent API endpoint

Calls /ai/analyze-scene but endpoint is never defined. Falls back to mock data, so app appears to work while core functionality is fake.

The industry knows it's broken

45%

of AI-generated code fails security review

2,500%

projected defect increase by 2028

1 in 5

organizations breached via AI-generated code

Aug 2, 2026

EU AI Act Article 50 deadline