The feature was rolled out after Anthropic did a “model welfare assessment” where Claude showed a clear preference for avoiding harmful interactions. When presented with scenarios involving dangerous ...