Anthropic accuses Alibaba of mass distillation attack on Claude models

This story crystallises an emerging fault line: as training frontier models becomes more expensive, the incentive to shortcut via distillation grow stronger, especially for actors constrained by export controls or capital. Anthropic claims that Alibaba‑linked operators ran nearly 29 million interactions through Claude via tens of thousands of fake accounts to extract capabilities and distill them into rival systems, arguing that this violates its terms of use and undermines its decision to keep products out of China. ([business-standard.com](https://www.business-standard.com/amp/technology/tech-news/anthropic-accuses-alibaba-of-illicitly-accessng-claude-ai-model-126062500120_1.html?utm_source=openai))

For the race to AGI, widespread industrial‑scale distillation has ambiguous effects. On the one hand, it accelerates capability diffusion: if Chinese labs can cheaply piggyback on US frontier models, they can close performance gaps without equivalent training spend. On the other, it may provoke a regulatory backlash—tighter export controls, more aggressive geo‑blocking, and cross‑lab sharing of abuse indicators—which could fragment the global model ecosystem. Anthropic, OpenAI and Google are already coordinating on detecting such attacks; Anthropic is now explicitly asking Washington for clearer antitrust safe harbors and penalties for firms that abuse distillation. ([business-standard.com](https://www.business-standard.com/amp/technology/tech-news/anthropic-accuses-alibaba-of-illicitly-accessng-claude-ai-model-126062500120_1.html?utm_source=openai)) How regulators respond will shape whether AI remains a partly open, API‑mediated commons or hardens into balkanised blocs with heavily policed boundaries.