Distillation Attack

A distillation attack is a technique where someone repeatedly queries a commercial AI model and uses the outputs to train a separate model that imitates its behavior, often without permission and in violation of the provider's terms of service. For enterprises, this matters because it can expose proprietary capabilities and create intellectual property disputes, and it raises questions about whether employees or vendors are generating training data from third-party models. Compliance teams should review API usage policies, monitor for bulk querying, and check contracts for restrictions on using model outputs to build competing systems.

1 item