By proceeding with a VAR in levels, you would miss a transparent representation of the system due to not having explicit error-correction terms as in a VEC model, but the coefficient estimates should be consistent.
By proceeding with a VAR in first differences, you would have two over-differenced series (with the resulting integrated MA(1) patterns) and missing error-correction terms (omitted variables).
What you should do is build a VEC model with the five cointegrated series and add the two stationary series at levels (both as dependent and independent variables, thus two extra equations and additional lagged terms in each equation corresponding to the two series).
See also "VAR or VECM for a mix of stationary and nonstationary variables".