Anthropic makes it harder for rivals to copy Claude AI
What's the story
Anthropic has announced that its latest security measures have drastically cut down data theft in so-called distillation attacks by nearly 90% to 99%. The company is a major player in the AI industry and has been working hard to protect its Claude AI models from being replicated by competitors. Model distillation is a major concern for top AI companies amid the growing US-China rivalry.
Industry challenge
What is model distillation and why it matters?
Leading AI firms like Anthropic, OpenAI, and Google have raised alarms over model distillation. They argue that this practice allows competitors to replicate sophisticated capabilities faster and cheaper.
To combat this threat, Anthropic has implemented safeguards limiting access to Claude's reasoning.
The company claims these measures have drastically reduced extraction attempts by making "the number of reasoning tokens extracted by distillers" 10 to 100 times lower than it was before.
Security enhancements
False positive rate of less than 0.1%
Anthropic has also made it harder to use proxy networks for probing Claude's protected reasoning.
This move is aimed at blocking extraction attempts and thwarting large-scale distillation efforts.
The company claims its safety system is "roughly as robust as those that prevent Claude from helping build bioweapons" and has "a false positive rate of less than 0.1%."
Account bans
Zhipu flagged as example of distillation attempts
Anthropic has banned tens of thousands of accounts tied to distillation campaigns.
The company also flagged Chinese AI start-up Z.ai, or Zhipu, as an example of a firm that tried to distill capabilities from Claude while building its GLM-5.3 model.
Anthropic warned such practices pose safety and national security risks.