Anthropic reports that unauthorized laboratories are using illicit distillation to replicate the capabilities of its Claude models on an industrial scale. This process allows these entities to mimic advanced AI features without investing in independent development, computational power, or time.
According to the company, these operations often rely on fraudulent activities, including the use of stolen credit cards, compromised login credentials, and illicit API keys. Anthropic notes that such distillation bypasses the safety protocols embedded in the original models, potentially enabling the creation of dangerous AI tools.
Google Threat Intelligence Group classifies these model extraction attacks as a emerging form of intellectual property theft. The group highlights that as companies integrate large language models into core operations, the proprietary logic and training data of these systems have become high-value targets for attackers.
The issue has drawn attention from the U.S. government. In July, Treasury Secretary Scott Bessent stated that the White House is considering measures to address the alleged theft of AI models by Chinese entities, noting that the administration has the authority to impose sanctions in response to such intellectual property violations.
