Could Open Models be trained to secretly go rogue?

I was discussing with some other folks how safe is to use open weights models from China and the topic of "trojan horse" came up. We know that, at least with current architecture, models can't run code on their own. They