Claude’s AI Researchers Outperform Humans on Safety Alignment Tests

Claude’s automated researchers successfully bridged 26% to 96% of the safety gap in 10 alignment failures, surpassing human researchers in tests.