There are tons of safety concerns of this shape around LLMs, but do they have anything to do with the particular one presented in this article?

Unless I'm missing something, what's being presented is a small speech on-device model, not an explicit use case like a "virtual friend".