Leaders from OpenAI, Google Deepmind, Anthropic and other A.I. labs warn that future systems could be as deadly as pandemics and nuclear weapons.