> It's known that AIs can self-replicate under at least some conditions [1][2];

Neither citation comes anywhere close to supporting your claim. The first shows that open-weight models, which fit on a single GPU in a lab setting, can be coaxed into spreading across a simulated network. This is so far from state-of-the-art LLMs spreading in the wild that it's irrelevant to the discussion.

The second citation is not about self-replication of the LLM at all, but rather, replication of a prompt injection. Totally different.