Think of like calling someone's workplace asking when their next shift is. There are a lot of good reasons someone might want to know this that the worker would be ok with, delivering a gift or whatever. Or it could be someone intending to harm them. We've mostly figured it's not worth having the person on the phone try to figure out in the moment if this is genuinely bad or not, so they follow the "safeguard" of just not giving the information.
The problem here isn't really that safeguards exist on language models. They pretty much have to to some extent for most uses. The problem is more in who controls the policy.