Anthropic reported on Thursday that in the past few months, there has been a continuous increase in distillation attacks targeting its models. The company stated that these activities attempt to bypass protections in order to extract the Claude's reasoning, tool invocation, programming, and data analysis capabilities.

The scale of the attack has expanded.

According to Anthropic, the related activities have expanded to nearly 200 million interactions, involving 5 separate campaigns. The company also stated that these attacks are more significant and more organized than the cases disclosed in February of this year.

Alibaba-related activities are the most concentrated.

Anthropic believes that the largest-scale activity is related to Alibaba. The company stated that a total of 151 million interactions were observed between May and July, with a peak of nearly 3 million interactions per day, distributed across 3,500 accounts.

Moonshot AI and DeepSeek are also mentioned.

The report also mentioned that activities related to Moonshot AI seem to be directly transferred from the Chinese military. Anthropic stated that another request asked Claude to evaluate closed-circuit surveillance footage to determine if the target was "behaving abnormally." The company has previously attributed similar activities to DeepSeek.