Anthropic has detailed its safety strategy to try and keep its popular AI model, Claude, helpful while avoiding perpetuating harms.