A recent evaluation tested three frontier robot policies against five unsafe instructions to assess their overall safety compliance. The study measures how often these AI models successfully refuse harmful commands compared to how frequently they execute them.