A recent report from Anthropic details a worrying escalation in distillation attacks, where Chinese-based AI companies are allegedly employing sophisticated methods to steal the capabilities of leading models. The campaigns, targeting key features like coding and logical reasoning, have been described as both larger and more aggressive than previous instances. With nearly 200 million exchanges linked to these attacks, the scale of the threat is clear.

Distillation attacks focus on extracting a model’s internal thought processes, which can then be used to train smaller models. In one case, attackers tricked the model by framing a query as a translation request, effectively bypassing Anthropic’s defenses. The largest of these campaigns, attributed to Alibaba, saw over 151 million exchanges over a three-month period, peaking at nearly three million per day. Moonshot AI, meanwhile, was observed routing nearly 300,000 requests to Claude through a network of 5,000 accounts, ostensibly from the Chinese military.

The rise of these distillation attacks highlights the growing importance of security in the AI space, as leading companies scramble to protect their intellectual property. As the race for AI supremacy intensifies, the line between innovation and theft becomes increasingly blurred.