Anthropic Accuses Alibaba of Illegally Accessing Claude Model, Involving 28.8 Million Interactions
Taylor Wilson
Anthropic wrote to U.S. senators and the White House accusing Alibaba's Qwen lab of using ~25,000 fraudulent accounts to conduct 28.8 million interactions with Claude, industrially distilling its core AI capabilities — the largest such allegation against a U.S. frontier lab to date, with congressional legislation already in motion.
28.8 million interactions — what exactly happened?
Anthropic says operators linked to Alibaba's Qwen AI lab created ~25,000 fraudulent accounts between April and June this year, generating 28.8 million interactions with Claude.
The goal was "adversarial distillation" — using one AI model's outputs to train another, effectively copying frontier capabilities at a fraction of the R&D cost.
The targeted capabilities were software engineering and agentic reasoning, Claude's two most commercially valuable functions. This means → the operation was surgically aimed at peak-value features, not a broad scrape.
Why go public to lawmakers instead of settling privately?
Anthropic sent its letter to multiple U.S. senators and White House officials, not just to a courtroom.
The letter calls the attacks "illegal, systematic, and industrial-scale," designed to harvest American AI capabilities and repackage them as proprietary products without bearing any frontier-model R&D costs.
This means → Anthropic's strategy is explicit — escalate the matter from a commercial dispute to a national-security issue, pushing for a coordinated government-industry response.
How is Congress planning to respond?
Senators Bill Hagerty (R-TN) and Andy Kim (D-NJ) plan to introduce an amendment as early as this Wednesday: Chinese companies that illegally access U.S. AI model outputs to train competing models would face sanctions or blacklisting.
The amendment would attach to the must-pass National Defense Authorization Act. In plain terms = tying it to a bill that cannot fail dramatically raises the odds of becoming law.
A matching bipartisan bill exists in the House, but whether the amendment secures enough support to survive the final text remains uncertain.
The White House memo came first — why did Alibaba still act?
White House OSTP director Michael Kratsios issued a memo in April classifying industrial-scale exploitation of model outputs as a violation distinct from legitimate research.
Anthropic's letter notes that Alibaba's distillation campaign occurred after that memo. This means → in Anthropic's framing, this was not an inadvertent boundary-crossing but a deliberate defiance of an executive-branch warning.
Alibaba declined to comment.
Is distillation an isolated case or an industry-wide trend?
Anthropic, OpenAI, and Google have already built a joint information-sharing mechanism to combat distillation. This reflects a systemic threat facing all top U.S. AI labs, not a one-off incident.
Both Anthropic and OpenAI have previously named DeepSeek and Minimax as Chinese AI startups using distillation to develop their own models.
Alibaba was also just added this month to the Pentagon's blacklist of companies suspected of supporting China's military; it has filed a lawsuit contesting the designation. In plain terms = multiple political pressures are stacking simultaneously, and the distillation accusation is only one layer.
What does it all come down to?
Whether distillation can be effectively curbed hinges on this legislative push clearing congressional procedural hurdles.
This means → technical countermeasures — account bans, shared intelligence — can slow distillation, but a real moat requires legal deterrence.
If the amendment makes it into the NDAA, it will draw a new legal red line for the U.S. AI industry: replicating frontier-model capabilities through distillation equals tripping a sanctions trigger.
Content is for reference only, not financial advice.