I think there is also an unlikely but reasonable concern that there could be exploits intentionally trained into the model itself.

https://news.ycombinator.com/item?id=50157302