Two ideas for projects/exercises, which I think could be very instructive and build solid instincts about AI safety:

1. Builder-breaker arguments, a la ELK
2. Writing up a safety case (and doing the work to generate the underlying evidence for it)