Comment by gck1

That's brilliant, I should try that.

I usually just start by preloadig context with plausible legitimate use, have it work and obviously fail, and then ask to figure it out without ever mentioning any high risk words. Model offers to RE itself and classifiers are happy.