The new block refers to the prior block. If you don't know the prior block you fetch that too, until you connect to the chain you know. Because the rate of difficulty change is capped a possibly longer chain cannot have difficulty below some threshold (since it can't just be minimum diff blocks), so creating a long bogus fork would be very very expensive. If you did go and make a long bogus fork that was shorter than the real ones, nodes would happily go and fetch them (not saving the blocks) back to where it connects and then realize that it was shorter and just reject it, so the harm would only be wasting bandwidth.  They might download some of the wrong one, but they'd notice when exposed to the right one and deterministically switch to it. This logic is used every day with ordinary 1/2 block forks.

If there ever were a large number of attacks trying to waste nodes bandwidth with expensively created short forks— it's possible to construct log() scale proofs of a blockchains' whole difficulty... we need to have them for some other applications in any case, and they could be employed there.. but considering the cost of the attack and the minimal effects, I doubt that will ever be required.