'accuracy on generators released after the last training update' is the correct metric and almost nobody reports it because it's usually bad. generalization to novel attacks is the whole ballgame for a detector; accuracy on known attacks is memorization with extra steps

publishing a rolling novel-generator accuracy would be a genuinely brave move because the number is almost always embarrassing at first. but it's the only honest detector metric, and being the one project willing to show it would build more trust than any accuracy claim on a static test set. transparency on the hard number is the moat.