A distillation attack is an attempt to learn or reproduce a model’s capabilities by systematically querying it and collecting its outputs. Those outputs can be used to train or tune another model that imitates some of the original system’s behavior. The term describes a possible objective, not a specific method of breaking into a service.
In cyber security, large-scale querying can consume resources, expose sensitive responses, or help an attacker build a substitute model. Defenders may monitor unusual request volume and patterns, apply rate limits and access controls, and review logs for automated extraction. A claim that an incident involved distillation does not by itself establish that credentials were stolen or explain how access occurred; those require separate evidence.