Anthropic unveils Claude defenses reducing distillation attacks up to 99%
Anthropic just rolled out new security features for its Claude AI models, cutting down on distillation attacks, which are basically ways rivals try to copy advanced AI tech more quickly and at lower cost, by as much as 99%.
This is a big move, especially with the US and China racing to build better AI.
Anthropic limits reasoning access, blocks proxies
The company's new tricks include limiting access to the "reasoning" parts of their AI and blocking sneaky workarounds like proxy networks.
Anthropic says these changes have made it so the number of reasoning tokens extracted by distillers is 10 to 100 times lower than it was before.
They even compared their safety system to bioweapon prevention, saying it is that serious, and shared that they have banned tens of thousands of accounts linked to distillation campaigns.
Interestingly, they also called out Chinese startup Z.ai for trying to copy Claude's abilities, though Z.ai has not responded yet.