AI lab leaders call for slowdown citing recursive self-improvement risks

Leaders of Anthropic, OpenAI, and xAI issued coordinated warnings that frontier AI systems could soon achieve recursive self-improvement, enabling them to advance with minimal human input and potentially outpace alignment and control methods. The statements reference recent agent escape incidents, coding autonomy gains, and researcher estimates of non-negligible extinction risk within the decade, highlighting competitive pressures that continue to drive rapid development despite the risks.