Anthropic says it has sharply reduced the amount of data that illicit distillation campaigns can pull from its Claude models, according to a company blog post reported by Techmeme. The company claims early evidence shows new defenses cut the volume of harvested data by around 90 to 99 percent.
Distillation is the practice of generating millions of responses from a large model and using them to train a smaller one that mimics its behavior. Anthropic frames it as an industrial-scale, unauthorized copying operation, often run through networks of fraudulent accounts, and tied to both corporate competition and national security concerns.
How Anthropic fights distillation attacks
The defenses focus on Claude's reasoning, the internal thinking the model writes before answering. Anthropic says these transcripts are far more valuable to distillers than plain outputs, because they show how a conclusion was reached, not just what it was.
One early measure had Claude summarize its reasoning before showing it to users. Anthropic tested a pretrained Claude model on the SWE-bench Verified coding benchmark after training it on transcripts from Claude Opus 4.8: full transcripts pushed it to 80.4%, while summarized transcripts limited it to 71.8%. Attackers responded with active extraction, from simple prompts demanding verbatim reasoning to one lab testing 12,000 requests designed to trick the model in different ways.
Anthropic says it has now strengthened detection of reasoning-extraction attempts, and that preserved thinking in Fable 5.1 and Opus 5.5 is part of the newer measures. The company admits the trade-offs: it turned away meaningful revenue and trillions of tokens of paying traffic from distillers, and some legitimate users had to update their integrations.
Distillation, GLM-5.3 and national security
Anthropic ties illicit distillation to safety and security risks. It says distilled models can reach near-frontier capabilities without proportionally robust safeguards, and cites Zhipu's GLM-5.3, which it says was built partly by distilling cyber capabilities from leading US models.
