Comment by skybrian

Yes, that's brought up in the first part of the article. She goes on to discuss differing performance depending on the language being used and its effect on safety guards. Apparently some language models do quite a bit worse in some languages. (The language models tested aren't the latest ones.)