Anthropic said Claude improved safety scores across 10 alignment failures and trained smaller models without reducing their general capabilities in tests.