Anthropic has warned that its models are nearing the threshold for recursive self-improvement. Its safety policy will retain human supervision as AI increasingly contributes to developing newer models.