I hadn’t thought about testing the bounds of model safety on comparatively benign requests compared to the type of thing described in frontier model cards.
Edit: also I've officially had AI damage hardware with that "wrong format wav" trick. $0.70 speaker was kaput.