<p>Electroencephalography (EEG) foundation models are increasingly used as general-purpose backbones for brain-computer interfaces (BCIs) by leveraging large-scale pretraining and task-specific adaptation. This review summarizes recent progress in EEG foundation models from three perspectives: datasets and task coverage, with emphasis on how generalization goals are operationalized by split protocols and concrete evaluation procedures; model design choices, including input construction and tokenization, masked pretraining objectives, and Transformer backbones for spatiotemporal modeling across heterogeneous channel layouts; and downstream adaptation, comparing linear probing, full fine-tuning, and parameter-efficient tuning, while clarifying the conditions under which each setting is most informative. We emphasize that reported gains are often protocol-dependent, as differences in task scope, preprocessing, training budget, and baseline selection can substantially affect comparability and the extent to which conclusions generalize. Finally, we outline future directions for EEG foundation models in BCI, focusing on standardized evaluation infrastructure, EEG-tailored modeling choices, and deployment-aware adaptation under real-world constraints.</p>