Anthropic proposes verifiable training pauses and global coordination to manage recursive self-improvement risks in frontier AI · Digg