A biological foundation model is a large machine-learning model trained on biological data such as DNA sequences, gene expression profiles, protein structures, cell images, or experimental assay results. Like other foundation models, it learns broad patterns from massive datasets and can then be adapted to specific tasks such as variant prediction, drug-response modeling, or virtual-cell simulation.
In cyber security, these models matter because the data they consume and produce is highly sensitive and valuable. Attackers may try to poison training data, steal model weights, or extract confidential research data from outputs. Bias or poor validation can also lead to unsafe predictions if the model is used in clinical or biotech workflows. Defenders reduce these risks with access controls, dataset provenance checks, secure training pipelines, audit logging, and privacy techniques such as differential privacy or federated learning. The core security problem is not only whether the model is accurate, but whether it is trustworthy, resilient, and protected against manipulation.