Anthropic Reports Chinese Labs Clone Claude Capabilities
Seven labs based in China have targeted Anthropic’s generally available Claude models in illicit distillation attacks, Anthropic said in a report released Thursday (Sept. 10).
Unauthorized labs use illicit distillation to extract and mimic capabilities from frontier models, without having to invest the time, computational power and cost it would take to develop them independently, according to the report.
While distillation can be a legitimate training method, it is illicit when the user is not authorized to undertake the process, per the report.
“We define illicit distillation as an industrial-scale, covert campaign to extract a model’s capabilities and replicate them in another model without authorization,” Anthropic said in the report. “Illicit distillation is typically enabled by fraud: sophisticated networks of fake accounts created with stolen credit cards, login credentials and API keys.”
We’d love to be your preferred source for news.
Please add us to your preferred sources list so our news, data and interviews show up in your feed. Thanks!
Anthropic has been developing more effective methods to combat illicit distillation even as unauthorized labs develop new techniques to bypass hurdles and carry out these attacks, according to the report.
Concerns around illicit distillation include the fact that safeguards built into Claude do not transfer when models are distilled by an unauthorized lab, that models distilled from a frontier model can help achieve dangerous capabilities and that exchanges relayed from users of third-party model routing services may contain users’ sensitive data, per the report.
“As we investigate and disrupt distillation attacks, what we learn will continue to inform the safeguards we build,” Anthropic said in the report.
Google Threat Intelligence Group (GTIG) said in a February blog post that “distillation attacks” or “model extraction attacks” are a new form of intellectual property theft that arose around the growing adoption of AI models.
“As organizations increasingly integrate LLMs [large language models] into their core operations, the proprietary logic and specialized training of these models have emerged as high-value targets,” GTIG said.
Treasury Secretary Scott Bessent told Fox Business in July that the White House was weighing a crackdown on China’s alleged “IP theft“ of AI models in the United States.
“If we see, especially that overseas models are stealing from our great companies, we have the ability to sanction them because of this theft,” Bessent said.
