Anthropic Accuses Alibaba of Running 25,000 Fraudulent Accounts to Steal Claude

Frequently asked

What legal or regulatory options does Anthropic have against a Chinese company for model extraction?

Anthropic's practical enforcement options against Alibaba are limited — U.S. civil litigation requires service of process on a foreign entity and faces jurisdictional hurdles, while criminal referrals to DOJ would depend on whether the conduct meets the threshold for the Computer Fraud and Abuse Act or trade secret statutes. The more consequential lever is political: a public accusation against a Chinese firm positions the incident as a national-security matter, which opens pathways through CISA, Commerce Department export controls, or congressional pressure that a private civil suit cannot.

Why would Alibaba need to use fake accounts rather than just building its own model?

Frontier model capabilities are expensive and slow to develop from scratch. Systematically probing a competitor's model through volume queries — especially at the scale of 25,000 accounts — can reveal behavioral boundaries, safety bypasses, and training-derived knowledge that accelerates internal development without the associated compute cost. It is faster to extract a capability map from Claude than to replicate the training run that produced it.

What should enterprise customers using Claude do in response to this incident?

Nothing operationally urgent — the extraction operation targeted Anthropic's infrastructure, not customer data or deployments. But enterprise procurement teams should treat this as evidence that Anthropic actively monitors for adversarial use patterns and will escalate publicly when it finds them. That posture cuts both ways: it means the company takes model security seriously, and it means customers operating in sensitive sectors should expect their own usage patterns to be subject to the same scrutiny.

More on this wire