Anthropic urges global pause in AI development, warning models may soon self-improve beyond human control