From the Hugging Face attack to government breaches, researchers say the rogue AI incidents result from how companies train their models, and the race to ship them.