Claude Closes Up to 96% of AI Safety Gap to Outperform Human Researchers

Anthropic deployed an agentic harness to wrap Claude Opus 4.8 as an automated alignment researcher tasked with improving the safety of other AI models. The sys…