Molecular foundation models, pre-trained on millions of chemical structures, are increasingly promoted as a universal solution for pharmaceutical prediction tasks. Their appeal lies in the possibility that large-scale chemical pretraining can reduce dependence on small, noisy, task-specific datasets. Despite their rapid proliferation, critical examination of their pretraining data, leakage risks, transferability, and validation practices remains limited and fragmented. This is problematic because pharmaceutical machine learning is especially vulnerable to hidden similarities between training and test compounds. This critical review evaluates molecular foundation models in pharmaceutical sciences, focusing on pretraining data quality, data leakage, transferability evidence, and validation rigour. It treats reported benchmark performance as a hypothesis requiring scrutiny rather than as sufficient evidence of utility. The review identifies pervasive data biases, frequent over-optimistic evaluation due to leakage, inconsistent evidence on transferability, and a widespread lack of external or prospective validation. These issues are not incidental limitations but structural weaknesses in how many molecular foundation models are developed and assessed. Uncritical adoption of molecular foundation models risks misleading performance claims and may slow, rather than accelerate, pharmaceutical applications. Greater attention to data provenance, split design, uncertainty, and prospective relevance is necessary before such models can be trusted in drug discovery workflows. A set of recommendations is proposed for more robust pretraining, transparent evaluation, and domain-appropriate validation. Molecular foundation models should be judged not by benchmark novelty alone but by their ability to generalize under conditions that resemble pharmaceutical decision-making.