Anthropic wants to stop AI models from turning evil - here's how

Can a new approach to AI model training prevent systems from absorbing harmful data?