Anthropic Accuses Alibaba of Illicit Model Extraction in Letter to US Officials

Anthropic has accused Chinese technology giant Alibaba Group of conducting a coordinated campaign to illicitly extract the capabilities of its Claude AI model, according to a letter sent to US officials. The San Francisco based AI company alleges that Alibaba engaged in adversarial distillation, a technique used to replicate a proprietary model's behavior by probing it with a high volume of targeted queries.

From April through June, Anthropic says it detected approximately 28.8 million queries originating from about 25,000 accounts that it has linked to Alibaba. The scale of the operation, equivalent to roughly 315,000 queries per day, suggests a systematic effort to reconstruct Claude's underlying architecture and knowledge.

What Is Adversarial Distillation

Adversarial distillation is a method in which an attacker queries a target AI model repeatedly to reverse engineer its behavior. The attacker can then use the captured outputs to train a competing model without authorization. In this case, Anthropic asserts that Alibaba sought to obtain functional equivalents of Claude's capabilities for its own AI services, including its Tongyi Qianwen model family.

• The queries spanned Anthropic's Claude API endpoints, with traffic patterns designed to evade rate limits and abuse detection.
• Alibaba controlled accounts used multiple IP ranges, likely routed through VPNs and cloud proxies, to mask their origin.
• Anthropic identified the activity after noticing anomalous query distributions and response caching mismatches that matched known distillation fingerprints.
• Alibaba did not respond to requests for comment about the allegations as of press time.
• US officials have not publicly commented on the letter, which was sent to the Commerce Department and other agencies.

Implications for AI Security and Trade Policy

The accusation lands amid escalating US-China tensions over AI technology transfer and intellectual property theft. The Biden and subsequent administrations have tightened export controls on advanced AI chips and model weights, but the extraction of capabilities via API queries occupies a legal gray area. Adversarial distillation is not explicitly covered by most existing trade regulations, though companies typically prohibit it in their terms of service.

Anthropic's public complaint signals that frontier AI firms are now willing to name alleged state affiliated actors in an effort to force policy action. If US regulators determine that such extraction constitutes a national security threat, new rules could restrict cross border API access or require tougher vetting of foreign users. For now, the episode highlights how proprietary AI models remain vulnerable to large scale copying through their own public interfaces, a risk that grows as more companies offer API access to their most advanced systems.