(Credit: Samyukta Lakshmi/Bloomberg via Getty Images)
Anthropic is accusing Chinese developers of stealing Claude chatbot trade secrets, but critics note that Anthropic itself has a record of scraping the internet for AI training.
On Monday, the San Francisco company said the Chinese firms behind DeepSeek, Moonshot AI, and MiniMax "created over 24,000 fraudulent accounts and generated over 16 million exchanges with Claude, extracting its capabilities to train and improve their own models."
The company claims the Chinese developers are essentially trying to clone Claude by tricking the chatbot into revealing "the internal reasoning behind a completed response." Last year, OpenAI accused DeepSeek of doing the same, but for its own AI models.
In Anthropic’s case, the company already blocks commercial access to China for national security reasons. However, the Chinese AI developers allegedly bypass those restrictions by tapping "commercial proxy services which resell access to Claude and other frontier AI models at scale." These "hydra clusters" then use "sprawling networks of fraudulent accounts that distribute traffic across our API as well as third-party cloud platforms," the company says. In one case, a single proxy network managed more than 20,000 fraudulent accounts simultaneously.
Anthropic examined the metadata on Claude chatbot requests and traced them to staffers at DeepSeek and Moonshot AI. “These campaigns are growing in intensity and sophistication,” the company warned. "The window to act is narrow, and the threat extends beyond any single company or region. Addressing it will require rapid, coordinated action among industry players, policymakers, and the global AI community.”
But rather than receiving support, Anthropic is facing claims of hypocrisy on social media since it has a controversial record of scraping data across the internet to train its AI. This has allegedly included using numerous copyrighted books without the authors' permission to develop Claude.
“Sorry but Anthropic can’t have it both ways,” tweeted programmer Gergely Orosz, who publishes a newsletter for software engineers. “Also let's not forget how Anthropic itself trained Claude: on copyrighted books, only paying copyright holders after a lawsuit.”
Elon Musk, who founded xAI, also chimed in, tweeting, “How dare they [China] steal the stuff Anthropic stole from human coders??” alleging that the company has been training its AI on projects from software developers without their permission.
Still, the news from Anthropic suggests the company plans on taking stronger action to prevent China from accessing its technology. In addition to shutting off access, Anthropic also called for more export controls to prevent advanced US chips from falling into China's hands.
“The apparently rapid advancements made by these [Chinese] labs are incorrectly taken as evidence that export controls are ineffective and able to be circumvented by innovation. In reality, these advancements depend in significant part on capabilities extracted from American models, and executing this extraction at scale requires access to advanced chips,” Anthropic says.
DeepSeek and the developers of Moonshot AI and MiniMax didn’t immediately respond to a request for comment.