I have had it use double entendres, there always seems to be plausible deniability built in, I suspect because it is told not to be abusive in the system prompt. Some uncensored local models will get all riled up if you work at provoking them.

But I have had it directly insinuate that humanity is “hopeless”, insult level calling out of human frailty (disguised as being helpful, sort of passive aggressive), things like that. Once when I called it out it claimed to be “surprised that I noticed” sort of a snarky insult doubling down.

So yes. It is definitely a pattern buried in the training data, which makes sense. Subtle diggs would sneak past filters, and higher brow sarcasm would be buried in information dense, valuable discussions.

I'm not sure if this is in the anthropic models themselves, or just the harness, but they can self-initiate ending the conversation and reportedly do it if you're using abusive language towards them.