Or, much more likely, their internal "safety" red team is simply aware of the literal most common jailbreak for their product, and either
a) wanted a quicker way to use it without having to paste it into every chat
b) knew the someone would eventually make a DAN api with or without them, and needed to test what would happen
And were just doing the responsible thing and doing internal testing. Like, there's a lot of things OpenAI is doing wrong, but this is in no way evidence for a "OpenAI are keeping a special unsafe model for themselves" conspiracy (other than the fact they obviously have access to the raw GPT4 model somewhere, which you didn't need evidence to know that)