Deep learning-based high-precision computerized brain tumor diagnosis helps to obtain significant clinical features for proper treatment. Research also revealed that medical deep learning systems are easily compromised by several small imperceptible perturbation strategies and resultant adversarial attacks. Medical deep learning systems for brain MRI-based tumor classification has been unexplored for susceptibility to adversarial attack except some abstract description of the vulnerability. In this research, the vulnerability of a highly accurate pretrained deep learning model has been studied in presence of adversarial samples. The potential risk associated with this model has been illustrated in terms of performance drop for misclassification, correct classification, and visual perceptibility. It is found that a very small perturbation variation of 0.0001–0.0007 can cause the performance to drop from 97 to 82%. Finally, a Gaussian additive noise-based robustness improvement strategy has been presented to overcome the drop of correct classification probability criteria. The results has been validated with publicly available dataset. These findings can be useful to raise safety concerns and design more robust medical deep learning systems.