Anthropic’s Frontier Red Team suggests some AI models “may be months away from being capable enough to self-replicate.”