LATEST: Anthropic claims Claude can now autonomously improve AI alignment, finding successful fixes across 10 failure categories without hurting model performance.