Model weights are the learned numerical parameters inside an AI model. They encode most of what the model has learned during training and strongly shape its outputs, making them a high-value asset rather than just a file. If weights are copied, leaked, or altered, an attacker may be able to reproduce a proprietary model, bypass safety behavior, or subtly change how the system responds.
In cyber security, weights matter because they must be protected like other sensitive production secrets. Real-world defenses include strict access control, encryption at rest, tamper-evident storage, signed releases, and change management so only approved versions can be deployed. Attackers may target weights through insider abuse, stolen credentials, compromised storage, or supply-chain manipulation. Audits and integrity checks help verify that the weights in use are authentic and have not been modified, which is essential when organizations need to prove both security and model governance.