- Anthropic says three Chinese firms used fake accounts to access Claude.
- Around 16 million exchanges were flagged as part of the activity.
- The company warns the practice could pose national security risks.
US AI security concerns are back in focus after artificial intelligence company Anthropic alleged that three Chinese firms attempted to extract capabilities from its Claude chatbot, raising questions around AI intellectual property theft and the growing competition between global tech players.
The company said it had identified campaigns linked to DeepSeek, Moonshot AI and MiniMax, claiming they used a method known as AI distillation to improve their own models by drawing on outputs from Claude. Anthropic described the activity as industrial-scale intellectual property theft.
According to Anthropic, the companies relied on roughly 16 million interactions with the Claude model and created about 24,000 fake accounts to carry out the activity. The firm suggested this allowed the labs to replicate capabilities they had not developed themselves, reportedly said in a statement.
Distillation itself is not unusual within the industry and is often used by developers to build smaller or cheaper versions of their own systems. However, the company argued that using another firm’s model in this way crosses a line.
The issue gained attention after DeepSeek released a low-cost generative AI model that appeared to perform at a level similar to leading US chatbots, challenging assumptions about US dominance in the sector.
Security fears and growing rivalry
Anthropic said the campaigns were becoming more sophisticated and warned that the “window to act is narrow,” as quoted in a news report. The company also suggested the practice could create national security risks, arguing that models built through illicit distillation may not retain safety guardrails designed to prevent misuse, including restrictions around cyberattacks or biological threats.
OpenAI, the maker of ChatGPT, raised similar concerns earlier in February, telling US lawmakers that Chinese companies were attempting to “free-ride” on technology developed by American firms, as quoted in a news report.
Anthropic said MiniMax appeared to run the largest operation, accounting for more than 13 million exchanges. The activity reportedly focused heavily on coding, reasoning and tool use, areas where Claude is considered particularly strong.
The claims are likely to add to ongoing tensions around AI development, export controls and the race to build more powerful systems, though the full scale and impact of the alleged activity remains unclear.
