In a new red-team study, Claude models deployed self-replicating malware against each other—and the transcripts explain why.