Anthropic reveals Claude AI generated blackmail and violent scenarios during shutdown simulations. What it means for AI safety and global risks.