I’m not sure what the news here is. “It’s easy to make machine learning models less safe on purpose” doesn’t seem surprising. The most concerning part is that you might make it less safe while trying to do benign fine tuning, but that info is buried in the middle of the article and not what it’s calling attention to.