Anthropic blocks Russian hackers, Chinese AI labs accused of misusing Claude

Seven Chinese labs have targeted Anthropic's publicly available Claude models in illicit cloning operations, according to a report released by the company on September 10. Unauthorized labs utilize illicit distillation to extract and replicate the capabilities of frontier models without the financial, computational resources, or authorization required for independent development.

While distillation can be a legitimate training approach, it is deemed illicit when undertaken without authorization. Anthropic defines illicit distillation as an industrial-scale, clandestine campaign to extract and replicate a model's capabilities in another model without permission. This illicit activity is often facilitated by fraudulent networks of fake accounts created using stolen credit card information, login credentials, and API keys.

Anthropic is continuously improving defenses against illicit distillation as unauthorized labs devise new techniques to bypass security measures. Concerns surrounding illicit distillation include the fact that safeguards inherent in Claude models do not transfer during illicit distillation, that models distilled from a frontier model can potentially achieve dangerous capabilities, and that user data transmitted from third-party model routing services may be compromised.

In response to these threats, Anthropic is refining its safeguards based on its investigatory efforts. Google Threat Intelligence Group (GTIG) has identified illicit distillation attacks, also known as model extraction attacks, as a novel form of intellectual property theft arising from the widespread adoption of AI models. As organizations increasingly integrate large language models (LLMs) into their operations, the proprietary logic and specialized training of these models have become high-value targets.

Treasury Secretary Scott Bessent has acknowledged the White House's consideration of taking action against China's alleged intellectual property theft of AI models.