We make Vision-Language-Action models (VLAs) significantly safer (83.58% improvement, no performance loss) by explicitly integrating constraints via a new integrated safety approach (ISA) based on CMDPs/SafeRL, validated on our new benchmark.