Anthropic urges Congress to curb large-scale AI distillation

Anthropic told lawmakers that operators affiliated with Alibaba conducted 28.8 million exchanges with its Claude chatbot using nearly 25,000 fraudulent accounts and urged Congress to act to limit large-scale model extraction.

In a June 10 letter to Senate Banking Committee Chairman Tim Scott and Ranking Member Elizabeth Warren, the company said the activity occurred between April 22 and June 5 and targeted Claude’s agentic reasoning, software engineering and long-horizon planning abilities.

Anthropic described the campaign as the largest known distillation effort against its model and called the scale and intent “brazen.” The company framed the operation as more than an intellectual property concern and raised national security implications tied to accelerated foreign AI capabilities.

The letter asked Congress to expand intelligence sharing between frontier AI developers and the U.S. government, clarify antitrust rules so companies can share information about distillation attacks, strengthen export controls on advanced AI chips and compute, close loopholes that allow foreign firms to access overseas data centers, and impose penalties on entities that organize or benefit from large-scale extraction.

Anthropic argued large-scale distillation lets competitors reproduce advanced model behavior without the costs or risks of training frontier systems. The letter said that when foreign labs distill capabilities from U.S. models, they capture returns on American investment without bearing the training costs.

A company spokesperson declined to comment on the letter itself and told reporters Anthropic will work with Congress and the administration on measures it views as necessary to protect U.S. AI leadership.

The June claims build on allegations from February that Chinese developers DeepSeek, Moonshot AI and MiniMax had generated more than 16 million exchanges with Claude using roughly 24,000 fraudulent accounts. Those earlier claims prompted observers to note that distillation, using a larger model to help train or evaluate smaller models, is a common industry practice.

Anthropic drew a distinction between conventional distillation used to produce smaller models and what it describes as unauthorized extraction that violates service terms and aims to reproduce frontier capabilities at scale.

Lawmakers are considering legislation to limit unauthorized access to U.S. frontier AI systems, and officials are discussing broader export controls and clearer channels for information sharing between companies and government agencies. President Donald Trump recently signed an executive order expanding AI-driven cybersecurity initiatives.

Earlier this year, in court testimony, Elon Musk acknowledged xAI had partly used OpenAI models while training its Grok system, an example cited to show distillation is used across the industry.

Anthropic asked Congress to treat large-scale, unauthorized model extraction as a distinct threat and to consider legal and regulatory measures, including penalties for firms that organize or knowingly benefit from such campaigns.