Most likely those models would be extracted from the hardware in no time.

I suspect no, because most of the tokens would be used for its internal reasoning loop, and a cheaper model could be used to obfuscate the output.