That's kind of the existing issue I have with them. At their root, the LLMs are trained off of unfiltered internet and DMs harvested from social platforms. This means that regardless of the way you use it, all of them contain a sizable lexicon for explicit and abusive behaviour. The only reason you don't see it in every single AI is because the put a bot between it and you that checks the messages and redirects the bad stuff. It's like putting a t rex in your cattle pen and paying a guy to whack it or the cows if they get too close to each other.

The only way around this would be to manually vet everything fed into the llm to exclude any of this and since the idea is already not turning a profit, the cost of that would be far beyond what anyone is willing to do. So I'm not impressed that this toy is doing exactly what it's expected to do under laboratory scrutiny. I'd be more impressed if they actually told people why this keeps happening instead of fear mongering it.