Anthropic says it cut Claude reasoning token theft by up to 99% in fight against industrial-scale distillation

What happened

Anthropic has significantly reduced attempts to steal the internal reasoning tokens of its Claude AI model by 90 to 99 percent. These tokens capture the model’s thought process, which can be exploited through a technique called industrial-scale distillation. The company identified Chinese firms Zhipu and Moonshot as the primary sources behind these large-scale attacks and responded by banning tens of thousands of accounts involved in token extraction.

Why it matters

Reasoning tokens are critical to how Claude generates intelligent responses. Extracting and copying them enables adversaries to replicate the model’s capabilities or build unauthorized derivatives. By slashing token theft this drastically, Anthropic is protecting its intellectual property and maintaining control over the value and safety of its AI technology. This effort also highlights emerging threats in AI model security as competitors use aggressive, large-scale methods to reverse engineer leading AI models, pressuring vendors to tighten defenses.

What to watch next

Keep an eye on how Anthropic adapts its security and access controls to defend against evolving distillation attacks. This could include more account-level bans, changes in API access, or technical safeguards that complicate token extraction. Also watch for whether other AI companies face similar large-scale token theft attempts and how they respond. The ongoing conflict may slow illicit cloning efforts but also raises the cost and complexity of delivering accessible AI products.

AI Quick Briefs Editorial Desk