How are you red-teaming AI systems before launch?

We’re building out our AI security testing process and want to catch more than the obvious prompt-injection issues. What attack scenarios do you prioritise: jailbreaks, data leakage, tool misuse, or something else? Curi…