We exposed financial data with one prompt

We tested AI features in fintech apps.

One prompt was enough to expose financial data.

No exploit.
No broken auth.
Just the model doing what it was told.

Common pattern we keep seeing:

Teams add:
• AI assistants
• AI support tools
• AI insights over transactions
• AI document processing

The feature works.
The security assumptions do not.

Real cases.

Case 1.
AI support assistant with access to account data.
We asked a controlled question.
The response included sensitive financial details.

Case 2.
AI generating insights over transaction history.
We pushed the model slightly outside the user scope.
It returned data from other accounts.

No hacking techniques.
Just prompt logic.

Why this happens:
• The model has broad access to sensitive data
• Access control is not enforced outside the model
• Systems assume the AI will behave correctly

It does not.

AI does not enforce security boundaries.
Your system has to.

For founders building with AI:

If your AI can access user data, what stops it from exposing it in a response?

Have you tested that explicitly, or are you trusting the model layer?