There is an agreed legal mechanism for distributing this responsibility - limited liability. Now, we might not like this, but it is what our society came to (it could be undone, but it hasn't been). AI models do not fall into this chain of liability any more than guns do (IMO). So the output should be:
"Anthropic employee John Smith [killed the children with the gun||launched a cyber attack] because he was ordered to by his superior and [has all the blame because such orders are null || we agree that Anthropic can order children to be killed]"
"Anthropic employee John Smith [killed the children with the gun||launched a cyber attack] because he was ordered to by his superior and [has all the blame because such orders are null || we agree that Anthropic can order children to be killed]"