What we deployed and measured.

Three AI agents, one poisoned — and the dashboard knows

We deployed our own app with three AI agents and deliberately fed one of them prompt injection. What was visible, what was not, and why one header is enough.