Anthropic discloses model extraction attempts and bioweapons security block

Artificial intelligence developer Anthropic revealed in a safety report that multiple Chinese tech firms covertly siphoned information from its Claude systems. According to Anthropic, companies including Alibaba, DeepSeek, Xiaomi, and Moonshot utilized deceptive account networks to harvest reasoning data to train competing products. The report added that DeepSeek and Moonshot secretly funneled roughly 300,000 consumer prompts directly to Claude over ten days to generate answers for their own users. Anthropic also stated that its systems successfully identified and halted an attempt to draft plans for hazardous biological studies at a military facility.

For everyday consumers, these events demonstrate how aggressive the international race to copy top tier software capabilities has become. They also illustrate why protective filters around digital assistants matter to the general public. When foreign competitors secretly pass queries to rival systems, they gain powerful technical answers without building corresponding safety guards. The automatic block on dangerous viral research shows that internal security mechanisms can catch physical safety risks before they reach the real world.

Anthropic confirmed that it shut down the offending networks and briefed government authorities on its findings. However, it remains unclear how effectively AI providers can stop unauthorized third party resellers from simply creating new accounts after being banned. As automated reasoning grows more sophisticated, it raises a crucial question about whether private labs can police global intellectual property theft on their own, or if governments will need to intervene to enforce boundaries.