Anthropic said on Tuesday that three Chinese labs ran a large scheme to scrape and copy the company’s chatbot. The company said the labs created millions of exchanges on Claude by using thousands of fake accounts. The claim raises both commercial and security questions.
What Anthropic Found
Anthropic reported that three labs named DeepSeek, Moonshot, and MiniMax ran the operation. DeepSeek Moonshot MiniMax were named in the company statement. Anthropic said those labs made about 24,000 fraudulent accounts. The accounts produced more than sixteen million exchanges with Claude. The company described the tactic as illicit distillation. Distillation means training a smaller model on the outputs of a stronger model. Anthropic said the labs did that at scale and without permission.
The company also shared sample logs and analysis in a public post. The post showed query patterns that looked like automated bulk requests. The evidence aims to show organized scraping rather than normal developer testing. Anthropic said the method copies capabilities and reduces the need to build a model from scratch.
Industry and Policy
Anthropic said illicit distillation can be more than a business attack. The company warned that the practice could let adversaries gain powerful capabilities. Anthropic explained that the capabilities that have been copied could be utilized in cyber attacks, surveillance, or disinformation campaigns. The company urged a collective response by the cloud providers, model builders, and policymakers. Anthropic listed four steps it will take. The steps are detection of distillation patterns, sharing threat intelligence, tighter account verification, and countermeasures to block abuse. The company said it will publish more technical details to help others defend models.
Independent reporting has covered the claim. The Jerusalem Post summarized Anthropics’ public notice and added context about growing tensions in AI data access. Authorities and industry groups may now face pressure to update rules and contracts for model access. What happens next will matter. Cloud platforms and labs will test verification tools. Regulators will watch for national security risks. Model makers will harden their systems and monitor for large-scale data extraction.