What happened
Distillation, in plain terms, means getting a stronger model to show its reasoning at scale, then using those answers to train a smaller model. Anthropic says the targets here were Claude's agentic abilities and tool use, plus coding, data analysis and logical reasoning. In other words, the stuff customers pay for.
Anthropic describes the operators as unauthorized labs that, over several months, built increasingly clever ways around its defenses. Its stated goal for them: harvest the capabilities of US frontier models.
This is not the first time Anthropic has raised the alarm. It spoke up in February and even named specific labs back then. The new report, per Anthropic, describes campaigns that are both larger and more aggressive than those earlier ones.
OpenAI has also reported similar activity, which it pinned on DeepSeek specifically.
Why it matters
The recipe is simple. Anthropic says it normally keeps Claude's internal chain of thought hidden and shows users only summarized thinking. Distillation attacks go after that reasoning anyway, because it is the most useful training material a smaller model could ask for.
If Anthropic's account holds, rivals can shortcut years of expensive frontier research with a lot of API calls and some patience. That turns model access into a security problem, not just a billing one.
It also sharpens the US-China AI rivalry. Anthropic frames this as the extraction of American frontier capabilities, and it is publishing names. Expect the policy crowd to notice.
The allegations come from Anthropic's own report. The accused labs' side is not part of what we have, so treat this as a detailed claim from one party, not a verdict.
What to do with it
- Audit your API exposure.If you resell or wrap a frontier model, check who is hammering your endpoints and how.
- Watch the terms.Using one vendor's outputs to train a rival model is exactly what this fight is about, so read your provider's rules before you fine-tune.
- Ask vendors about defenses.Enterprise buyers should ask how their provider detects bulk extraction and what happens when it does.
- Track the named labs.If you use models from Alibaba, Moonshot AI or DeepSeek, follow how they respond to the report.
Keep up, sapiens. The models are copying each other's homework now, and you are still not invited to the study group.


