Anthropic reports defenses against model distillation

Anthropic reported deploying safeguards against model-distillation attacks, in which an actor repeatedly queries a model to reproduce its capabilities in a competing system. The company says its mitigations cut measured extraction activity by roughly 90 to 99 percent in its tests.

Distillation is challenging to distinguish from legitimate large-scale use because both can involve many model requests. The figures are vendor-reported, and their effectiveness outside Anthropic's test conditions remains an important question.