Anthropic accuses Chinese AI labs of stealing data from Claude

Investing.com -- Anthropic on Monday accused three Chinese artificial intelligence laboratories of conducting large-scale data extraction campaigns from its Claude AI model, alleging the companies used fraudulent accounts to illicitly improve their own models.

The AI company said DeepSeek, Moonshot, and MiniMax generated over 16 million exchanges with Claude through approximately 24,000 fraudulent accounts, violating terms of service and regional access restrictions. The allegations follow similar claims by OpenAI, which told House lawmakers that DeepSeek used distillation techniques to enhance its models.

The Chinese laboratories employed a method called distillation, which involves training a less capable model on outputs from a more advanced system. While distillation is a legitimate training technique used by AI companies to create smaller versions of their own models, Anthropic said the practice becomes illicit when competitors use it to acquire capabilities from other laboratories.

According to Anthropic, the campaigns targeted Claude's most advanced features, including reasoning capabilities, tool use, and coding functions. The company said it identified the operations through IP address correlation, request metadata, and infrastructure indicators.

DeepSeek reportedly generated over 150,000 exchanges focused on reasoning capabilities and rubric-based grading tasks. The operation used synchronized traffic across accounts with identical patterns and shared payment methods. Anthropic said DeepSeek's prompts asked Claude to articulate internal reasoning behind completed responses step by step, generating chain-of-thought training data. The company also observed tasks where Claude was used to create alternatives to politically sensitive queries about dissidents, party leaders, and authoritarianism.

Moonshot AI conducted over 3.4 million exchanges targeting reasoning, tool use, coding, and computer vision capabilities. The operation employed hundreds of fraudulent accounts across multiple access pathways. Anthropic attributed the campaign through request metadata matching public profiles of senior Moonshot staff.

MiniMax generated over 13 million exchanges focused on coding and tool use. Anthropic detected this campaign while it was active, before MiniMax released its trained model. When Anthropic released a new model during MiniMax's operation, the Chinese laboratory redirected nearly half its traffic within 24 hours to capture capabilities from the latest system.

Anthropic said that the laboratories accessed Claude through commercial proxy services that resell access to AI models. These services operate networks of fraudulent accounts that distribute traffic across Anthropic's API and third-party cloud platforms. In one case, a single proxy network managed more than 20,000 fraudulent accounts simultaneously.

Anthropic does not offer commercial access to Claude in China for national security reasons. The company said illicitly distilled models lack necessary safeguards, creating security risks. Anthropic and other US companies build systems that prevent actors from using AI to develop bioweapons or conduct malicious cyber activities.

Anthropic said it has built detection systems and behavioral fingerprinting tools to identify distillation attack patterns in API traffic. The company is sharing technical indicators with other AI laboratories, cloud providers, and authorities. It has also strengthened verification for educational accounts and security research programs

You May Also Be Interested In

• Jefferies Upgrades Shopify (SHOP) to Buy
• Nomura/Instinet Starts Malayan Cement Bhd (LMC:MK) at Buy
• B.Riley Starts Spruce Biosciences, Inc. (SPRB) at Buy