Ignoring concerns about missing integrity and authentication verification, here is why:
Suppose that the key you used to generate your keystream from your CSPRNG is shorter than your plaintext; if it isn't, you may as well use the one-time pad. Then I know from the pigeonhole principle that there will be some (many, actually) keystreams that will never be generated. Furthermore, I can enumerate these keystreams by simply going through every key (your plaintext has finite length, so there are only finitely many keys I have to check).
Therefore given any ciphertext, I can infer that there are many plaintexts that can never encrypt to this ciphertext, regardless of the key you picked. I have derived knowledge* about the plaintext from the ciphertext alone, therefore your encryption scheme is not information-theoretically secure (what you call "theoretically unbreakable").
Did I mention I was an entity with unbounded (infinite) computational power? If you want truly unbreakable encryption, this is effectively equivalent to your ciphertext conveying zero information about the corresponding plaintext without knowledge of the key. It's a very high bar to set, and one that is virtually useless in the real world, which is why we are perfectly okay with AES and the like.
You cannot cheat information theory by stretching your keys through deterministic processes. You cannot cheat information theory through elaborate encryption schemes involving multiple encryption passes with different keys and liberal sprinkling of hash functions and CSPRNG's - an unbounded adversary will see right through it, unraveling whatever layers of convoluted encryption you have going on, until it all eventually boils down to how much information your ciphertext has, which is itself fundamentally lower-bounded by how long your key is.
Information is given, not created, and your cipher's security is measured in terms of how much computation it takes to extract information out of the ciphertext without the key. If it carries no information, there is nothing to be extracted, and you have information-theoretical security.
* the length of the plaintext is not generally considered "knowledge" as it is more metadata than anything else; it cannot really be relied on as you can always pad your plaintext to a fixed length and communicate through a prearranged dictionary of fixed-length symbols.