It also gets confused if the entire prompt is in a text file attachment.

And the summarizer shows the safety classifier's thinking for a second before the model thinking, so every question starts off with "thinking about the ethics of this request".