Security researchers deliberately try to fool AI before bad actors do โ this is adversarial testing.