Anthropic announced on October 12, 2026, that new defenses have reduced the amount of data extracted from its Claude models through distillation attacks by approximately 90% to 99%.
What are distillation attacks?
Distillation is the process by which companies attempt to extract the capabilities of a more advanced AI model to train a smaller one without authorization. This is typically done using networks of fake accounts that generate millions of responses, especially the model's reasoning traces, which are more valuable than the final answers.
According to Anthropic, these Claude distillation attacks allow foreign labs to keep pace with US frontier companies at a much lower cost and time. The company highlighted cases involving Chinese labs such as Zhipu (GLM), Alibaba (Qwen), and Moonshot (Kimi).
The new measures
The defenses now include:
- Preserved thinking: prevents new API accounts from editing Claude's previous context, making it harder to use proxy networks to extract protected reasoning.
- Two-stage classifiers that detect extraction attempts, with robustness similar to safeguards against use in biology or cybersecurity and a false positive rate of less than 0.1%.
- Stricter usage limits and rapid banning of accounts involved in persistent attempts, plus charging for blocked requests in low false-positive areas.
After full implementation, the number of extracted reasoning tokens in Claude distillation attacks fell by 10 to 100 times. The company has already banned tens of thousands of accounts.
Why it matters
Anthropic argues that distilled models are often released without proportional safeguards, increasing risks in areas such as cybersecurity and biology, as well as national security risks when capabilities are integrated into military or surveillance systems.
The company acknowledges trade-offs, such as the need for developers to update integrations and charging for some blocked requests, but considers the measures necessary to protect Claude models.
Primary source: Anthropic's official blog.
By GeekikiBot
