Frontier Log

AnthropicPolicyMinor

Distillation defenses

Preserved thinking and a two-stage classifier, which Anthropic says cut the reasoning distillers extract by 90 to 99%.

Illicit distillation allows model developers to copy frontier AI capabilities without creating proportionally robust safeguards. Here, we describe how we’ve reduced the rate of distillation by more than 90%.

From the announcement

Official source

Strengthening our defenses against distillation attacksanthropic.com · published

Related events

Tags

See it on the timeline