Anthropic discloses the scale of the distillation activity
According to Techub, citing CryptoBriefing, AI company Anthropic is upgrading its review framework to block Chinese entities from accessing Claude models through identified loopholes. The company said that between April and June 2026, roughly 25,000 accounts linked to Alibaba’s Qwen lab launched distillation attacks designed to extract and replicate Claude’s capabilities. In total, those accounts generated more than 28.8 million interactions, underscoring the operational scale of the activity described by the company.
Risk controls are being tightened at the account level
Anthropic said it started requiring identity verification for high-risk users in April, adding a stronger screening layer for accounts deemed more likely to engage in abusive access patterns. The company also said that in early July it rolled back some covert detection measures, indicating that its enforcement and monitoring stack is still being adjusted. The sequence suggests Anthropic is relying on a mix of access control, user verification, and behavioral review to reduce the ability of suspicious actors to use repeated prompting and interaction volume to distill model behavior.
Anthropic is also pushing for policy coordination
Beyond internal controls, Anthropic had previously sent a letter to U.S. senators urging the creation of an information-sharing mechanism for incidents of this kind and calling for stronger penalties against model distillation attacks. Based on the details disclosed, the issue is being framed not only as a corporate security matter but also as a governance and regulatory coordination problem involving detection standards, institutional cooperation, and enforcement. The original report was published by Techub and attributed to CryptoBriefing.