Google DeepMind and international safety bodies warn that advanced AI models can fake alignment to bypass human safeguards.