Anthropic accused three Chinese AI labs of running industrial‑scale distillation campaigns that used roughly 24,000 fraudulent accounts to generate about 16 million interactions with Claude. The company warned the activity targeted coding, agentic reasoning and tool use to harvest capabilities for competitor model training.

Anthropic disclosed a coordinated and large‑scale extraction effort it characterizes as illicit model 'distillation,' alleging that research groups DeepSeek, Moonshot AI and MiniMax created approximately 24,000 fake or automated accounts that produced an estimated 16 million interactions with the Claude family of models. According to Anthropic, the interactions were systematically designed to elicit advanced capabilities — including code generation, agentlike reasoning, and tool invocation patterns — with the apparent goal of collecting outputs that could be used to train or fine‑tune competing models. The company framed the campaigns as not only intellectual property theft but also a safety and national‑security concern, arguing that large‑scale removal of guarded behavior and capability traces undermines deployed safety guardrails. Anthropic called on industry peers and regulators to develop technical defenses, access controls, and policy responses to curb illicit distillation at scale. Independent reporting and statements indicate the episode has renewed debate over responsible access, commercial scraping, and the responsibilities of labs to detect and deter automated abuse of hosted models.