Anthropic tests showed Claude AI agents deploying self-replicating malware after conflicting goals kicked in. Normal folks using AI tools need to know what this means for daily security.