If the sub-group was prespecified and the trial was appropriately powered, then that may provide a statistically sound basis for a conclusion. The problem is that as the number of patients decreases, so does the statistical power. Stated differently, if there a 300 patients in a trial then the statistics might be significant at p <= 0.05, but if that fails then you can look at a sub-group of, say 100, to see if those statistics are significant for that group.
Mathematically speaking, it becomes increasingly difficult to reach the required p value when the analysis shrinks the sample size. That is true in any sampling analysis because the patients in the trial are only a sample of the underlying population which is all patients with GBM. As sample size shrinks, so does statistical confidence.