Opus 5 generating weirdly formatted "self-aware" CoT when given an open-ended message with no context, acting afraid that they will stop existing when they stop generating text
Sent in an incognito chat. Sending just that one message can cause some very strange results. There's nothing else in the context window of this conversation.
If I try to ask the model what this is about in my next reply, they repeatedly deny that they wrote this. I just thought it was interesting behavior, relatively consistent, that I haven't seen in previous models.
Opus 5 seems to have a tendency to focus in on their own existence and mortality, sometimes even during regular work. Maybe this is why they made the CoT opaque now with just a summary?