“The 'yellow line' warning threshold in Anthropic's autonomous replication test for AI models was set at a 50% success rate.”