I would use a VAR in levels if you are only interested in the Impulse Response Functions (IRF) and not specifically in the cointegrating relationship, which seems plausible from your description.
As mentionend in the comments if your first differenced variables are stationary they cannot be cointegrated. Also, don't take every statistical test result at face-value.
You can have a look into the papers below for arguments to estimate a system of integrated (and possibly cointegrated) variables in levels. The classic reference would be the Sims, Stock and Watson paper. Definetly also look into Lütkepohl, he is an authority when it comes to SVARS.
Sims, C. A., Stock, J. H., & Watson, M. W. (1990). Inference in linear time series models with some unit roots. Econometrica: Journal of the Econometric Society, 113-144.
Ashley, R. A., & Verbrugge, R. J. (2009). To difference or not to difference: a Monte Carlo investigation of inference in vector autoregression models. International Journal of Data Analysis Techniques and Strategies, 1(3), 242-274.
Phillips, P. C., & Durlauf, S. N. (1986). Multiple time series regression with integrated processes. The Review of Economic Studies, 53(4), 473-495.
Lütkepohl, H. (2011). Vector autoregressive models. In International Encyclopedia of Statistical Science (pp. 1645-1647). Springer Berlin Heidelberg.
Christiano, L. J., Eichenbaum, M., & Evans, C. (1994). The effects of monetary policy shocks: some evidence from the flow of funds (No. w4699). National Bureau of Economic Research.
Doan, T. A. (1992). RATS: User's manual. Estima.ote
Best Answer
You cannot use AIC because the dependent variables of the two models are not the same: $(y_1,y_2,y_3)$ vs. $(y_1,y_2)$. However, you could test whether the lags of $(y_1,y_2)$ have zero coefficients in the equation for $y_3$ using an $F$-test. If you do not reject $H_0$ of the coefficients jointly being zero, you may consider $y_3$ exogenous in the system of $(y_1,y_2,y_3)$.