Anthropic has implemented new security measures designed to prevent competitors from replicating the capabilities of its Claude AI model. These protections focus on mitigating model distillation, a process where companies extract data to build cheaper and faster versions of advanced AI systems.
The company reports that these updates have reduced the volume of data stolen through distillation attacks by 90% to 99%. Specifically, the number of reasoning tokens extracted via these methods has dropped to between 1/100 and 1/10 of previous levels.
To achieve these results, Anthropic increased the difficulty for proxy networks attempting to probe Claude's reasoning processes. This approach effectively blocks large-scale extraction attempts by making repeated access to protected data significantly harder for unauthorized parties.
The security system now matches the robustness of the company's existing safeguards against the creation of biological weapons. Anthropic notes that the current implementation maintains a false positive rate of less than 0.1%, ensuring that legitimate user interactions remain unaffected.
