Anthropic details new defenses that cut Claude reasoning extraction by distillers
AIAnthropic says its new defenses against distillation attacks cut the reasoning data harvested by distillers by around 90 to 99%. The methods include controlling where reasoning can go, a two-stage classifier that blocks extraction attempts, and making repeated attempts harder through stricter limits and bans.
AIWhy it matters.
The post explains concrete defenses against distillation, with reported extraction reductions and the trade-offs they carry for legitimate users.
Source details
Anthropic Newsroom · anthropic.comPublished · added here
Original title: Strengthening our defenses against distillation attacks