U.S. moves toward voluntary AI model security tests: what labs should do