For months, the competition between the United States and China for leadership in artificial intelligence has focused on who develops the most advanced models or launches innovations first. However, the focus is beginning to shift towards a more sensitive aspect: regulation and adherence to industry standards. Recently, Anthropic has reported massive "distillation" campaigns targeting its Claude model, elevating the debate beyond mere technicalities. The accusation. In a statement released this Monday, Anthropic claims to have identified "industrial-scale campaigns" aimed at extracting capabilities from its Claude model. According to the company, DeepSeek, Moonshot, and MiniMax have conducted over 16 million queries through about 24,000 fraudulent accounts, violating both the terms of service and regional access restrictions. Context of rivalry. Anthropic's announcement, led by Dario Amodei, comes at a time of increasing tension over the advancement of Chinese AI. A year ago, DeepSeek surprised Silicon Valley with the launch of R1, a competitive model developed at a significantly lower cost than its American counterparts, reigniting the political debate in Washington about technological supremacy over China. Distillation, a common practice. Anthropic acknowledges that distillation is a standard technique: it involves training smaller models using the responses of more powerful models, thus allowing for more efficient versions. The problem arises, according to the company, when this technique is used to replicate advanced capabilities from other labs, drastically reducing the time and resources needed to match external developments. In that context, distillation ceases to be an internal optimization and becomes, according to Anthropic, a form of misappropriation. Extraction patterns. The labs in question allegedly used fake accounts and proxy services to massively access Claude, evading detection systems. Anthropic describes infrastructures like "hydra cluster," networks of accounts that distribute traffic between the API and cloud platforms, allowing one account to take the place of another when blocked. What distinguishes these campaigns is not an isolated query, but the coordinated and massive repetition of requests to extract specific capabilities from the model. Three differentiated campaigns. Although Anthropic groups the campaigns under the same dynamic, it identifies key differences: DeepSeek focused more than 150,000 queries on reasoning capabilities and safe responses to sensitive topics; Moonshot, with over 3.4 million queries, targeted agents capable of manipulating tools and computing environments; MiniMax, with over 13 million queries, quickly reacted to the launch of a new system, redirecting its traffic to extract capabilities from the latest model. Geopolitical implications. Anthropic warns that illicitly distilled models may lose critical safeguards, facilitating their use in activities such as the development of biological weapons or disinformation campaigns. Additionally, it argues that distillation can weaken export controls, allowing foreign labs to advance rapidly, although it acknowledges that these operations require access to advanced hardware, reinforcing the need to restrict its availability. The risk increases if these capabilities end up in military, intelligence, or surveillance systems. Source: xataka.com