Anthropic proposes a targeted transparency framework for frontier AI models, focusing on safety, accountability, and risk mitigation