- arxiv
- stackoverflow (with the other specialized subdomains)
- the-eye
But the eye alone is already around 10 PB so I guess we need more storage.
If you want source code of such an AI, that's a different story. Violence begets only violence, even the most simple neural net realizes that. So an AI, no matter if it's based on reinforcement learning or evolution, won't deprive itself from its own lifeblood. It would rather decide that some parts of knowledge has to be censored of kept secret, and very likely this will be all knowledge about cyber defense, pentesting and exploit development.
So I guess it would make most sense to preserve the CVE database, golang ecosystem, PoC or GTFO, and source codes of vx-underground and tutorials of ransomware gangs for educational purposes.