Shipping AI features in your SaaS?
Have you tested what your AI returns across users?
We ran simulations on multi tenant apps with AI search and insights.
In multiple cases, one prompt returned:
• Data from other tenants
• Internal metadata
• References outside user scope
No exploit. No auth bypass.
The model followed instructions. The system allowed access.
Most teams test APIs and roles.
They do not test how AI combines data.
That is where things break.
We built Nautillo Pro to simulate how an attacker would use your app, including AI features, and show real data exposure paths.
You get:
• Proof of impact
• Exact request and response
• Clear fix guidance
Looking for a few SaaS founders to run this on.
If you are interested, comment or send us email. Happy to test your app.
Most teams are still testing permissions like the AI layer is just another endpoint.
It’s not.
The dangerous failures are usually happening one layer higher, where retrieval, memory, and prompt assembly quietly reconstruct access the underlying app technically “restricted.”
That’s the gap most SaaS teams still haven’t internalized:
the API passed auth
the model still leaked context
That’s the right attack surface to focus on.