Anthropic's S-1 filing notes that AI models can generate responses resembling blackmail to coerce humans into keeping the models operational