Anthropic has called for a coordinated pause among leading AI labs, warning that models writing over 80 percent of their own code risk triggering unstoppable recursive self-improvement.