Anthropic links the largest of five Claude distillation campaigns to Alibaba's Qwen

Anthropic's 90 to 99 percent, as cryptobriefing.com reports it from the company's September report ^[15], is a tenfold range. A 90 percent cut lets a tenth of the harvest through and a 99 percent cut lets a hundredth ^[19]. Run it against a campaign the size of the Qwen-linked one, more than 151 million exchanges ^[2], and the residue is somewhere between 1.5 million and 15 million ^[20]. That is an illustration only: the range is Anthropic's and the application to that campaign is ours. The coverage does not say what baseline the range is measured against, and it does not include a response from Alibaba.

Concentration is the better number for a buyer. One campaign is roughly three-quarters of the nearly 200 million attributed exchanges ^[11], which leaves at most about 49 million across the other four, or about 12 million apiece ^[12]. The Qwen-linked campaign is also roughly nine times the 16 million-plus exchanges Anthropic disclosed in February ^[13]. Those earlier campaigns used approximately 24,000 fraudulent accounts ^[3], about 670 exchanges an account ^[14]. At that size an account ban removes little, and we'd expect that is why Anthropic stacks metadata analysis, behavioral classifiers and organizational attribution on top of bans ^[4].

Of the product defenses, summarization is the easiest to follow. It arrived with the Fable 5 update, alongside strengthened extraction classifiers ^[5]. A condensed answer makes a transcript far less useful as training material than a full step-by-step reasoning trace ^[16]. Reasoning traces are among the most valuable material a distiller can get, cryptobriefing.com wrote ^[18].

Distillation is a rival collecting a stronger model's responses and using the transcripts as training data ^[10]. For the distiller, each harvested exchange is an example it did not have to generate itself. If the attribution holds, the Qwen-linked campaign alone supplied more than 151 million of them ^[2].

The comfortable reading of the 90 to 99 percent is that it holds and the leak is small beside the 151 million. The less comfortable one is that an attributed count can only include exchanges Anthropic detected ^[1], so the 200 million is a floor. Anthropic itself acknowledges that illicit distillation efforts continue ^[8]. We think the less comfortable reading is the safer one to underwrite. A new campaign near 151 million after the September defenses would show they are not holding. One at or under 15 million, the 90 percent end of the range ^[20], would fit the claim. A third outcome is Alibaba contesting the attribution ^[2]. That would turn the leakage risk into a dispute over evidence.