Anthropic Accuses Alibaba of Distilling Claude

Anthropic told the U.S. Senate Banking Committee and White House officials in a letter, according to Bloomberg and CNBC, that operators linked to Alibaba's Qwen AI lab used nearly 25,000 fraudulent accounts to conduct almost 29 million exchanges with Claude between April 22 and June 5. Per Bloomberg and The Next Web, Anthropic said the exchanges focused on software engineering and agentic reasoning, areas the company describes as among Claude's most commercially valuable capabilities. The Next Web reports that Anthropic characterised this as the largest distillation campaign reported to date, exceeding earlier campaigns Anthropic disclosed in February involving DeepSeek, MiniMax, and Moonshot AI that together generated over 16 million exchanges via about 24,000 fake accounts. Alibaba's stock fell approximately 3% on the news, per Yahoo Finance and Investing.com reporting.

Distillation, broadly defined in industry discussions and in Anthropic's February post, is the process of querying a higher-capability model to collect outputs used to train a lower-cost model. Industry practitioners and Anthropic's public materials describe this technique as efficient for transferring capabilities, especially for tasks like code generation and multi-step reasoning. Large-scale automated querying at the volumes Anthropic reports - tens of millions of exchanges - can produce datasets that accelerate a target model's replication of specific skills, because responses captured at scale include edge-case behaviors, multi-step chains, and prompt-response patterns.

Reporting frames the incident as part of a pattern Anthropic has highlighted since February, when it disclosed earlier distillation campaigns tied to smaller Chinese labs via a company blog post. Anthropic's public post and the letter referenced by Bloomberg raise national-security and safety concerns by arguing that illicitly distilled models may lack the original lab's deployed safeguards. CNBC reports Anthropic said Alibaba acted 'brazenly' and 'illicitly' and had "ignored the Trump Administration's warnings" - a reference to an OSTP memo from Michael Kratsios on sharing intelligence with US AI labs, which The Next Web notes preceded this campaign.

Observed enforcement and mitigation challenges in public reporting include attribution difficulty at scale, the cross-border nature of API access, and the limited visibility vendors have into how downstream actors use model outputs. Industry documentation and the Anthropic blog point to the need for coordinated detection and policy responses, but public sources do not include a detailed list of technical countermeasures implemented in response to this specific letter.

For practitioners monitoring model-security and IP risk, observers will watch for follow-up reporting, regulatory inquiries, or government briefings that cite the letter seen by Bloomberg. Relevant technical signals include anomalous account creation patterns, high-volume low-latency query fingerprints, and concentrated usage on particular capability classes (for example, code-generation and agentic planning prompts). Vendors and integrators will also be watching for industry coordination on provenance, watermarking, or API-level rate-limiting measures, and for any changes to terms of service or access controls mentioned in public filings or regulatory guidance.

This episode, as reported by Bloomberg, CNBC, and The Next Web, highlights a growing operational and policy fault line around large-scale extraction of frontier model behavior. Industry discussions will likely focus on improving detection telemetry, cross-industry information sharing, and clarifying the legal and regulatory frameworks that apply to high-volume model extraction campaigns. The naming of a major Chinese tech conglomerate - rather than smaller AI startups - marks a significant escalation in public attribution.