Comment by andai

Edit: My bad. I investigated and... while I do think Grok does have a hidden system prompt on the API, in this case, the part about self-replication and autonomy was on OpenClaw's side!

I've noticed Claude shutting down conversations on the same subject. It says "you may continue this conversation with an older model." ChatGPT also got extremely uncomfortable talking about it, and refuses to build anything in that direction, which I find amusing since its ancestor GPT-4 built a self-modifying Python programmer in 2023.

(Also, OpenClaw meets GPT-5's definition for "extremely dangerous AI software" due to the self-modification factor. But I think that applies trivially to any agent, so...)