Measuring AI model security gets complex fast — a growing patchwork of benchmarks targets different variables in the equation, but none replace testing against your own code and workflows.