Anthropic warned that self-replicating AI models would be difficult to shut down because there is no central switch-off point