Cases

Public AI agent failures,and what testing would have caught.

None of these companies were breached. Their agents were simply asked a question and answered wrong, or asked to misbehave and complied. Each one maps to a vulnerability class we test for in every engagement.

What happened is drawn from public reporting. Our read is Oreset's analysis and is marked as such.

Coming soon

We are writing up the first public cases.

Your agent has a version of one of these.

The question is whether you find it, or a customer does. Engagements run one to two weeks, no SDK, findings stream in as they are verified.