A model checkpoint is a saved snapshot of a machine-learning model at a specific point in training or deployment. It usually contains learned parameters and sometimes optimizer state, making it possible to resume training, roll back to a known good version, or restore a service quickly after failure.
In cyber security, checkpoints matter because they are high-value and often hard to recreate. Attackers may encrypt, delete, or corrupt them to disrupt AI services, force costly retraining, or undermine the integrity of deployed models. In ransomware cases, checkpoints are targeted alongside training data and vector databases because losing them can stop inference systems and delay recovery. Defenders protect checkpoints with access control, versioned backups, offline copies, integrity checks, and separation from the main application environment.