U.S.-based frontier AI lab Anthropic has escalated the geopolitical tech war by formally accusing Chinese tech and e-commerce giant Alibaba of executing a massive, industrial-scale campaign to illicitly extract capabilities from its Claude models.
In a newly revealed June 10, 2026 letter addressed to the U.S. Senate Banking Committee and White House officials, Anthropic detailed what it calls “the largest known distillation attack to date.” The accusation sent Alibaba’s shares tumbling over 3% to 4% across global exchanges.
1. Anatomy of the 25,000-Account Attack
Because Anthropic explicitly blocks access to its AI services for entities within China, the startup claims Alibaba deployed an extensive operational framework to systematically siphon its data:
- The Scale: Anthropic alleges that operators affiliated with Alibaba and its flagship Qwen AI lab created approximately 25,000 fraudulent proxy accounts to bypass geographic locks.
- The Volume: Between April 22 and June 5, 2026, these fake profiles generated an astonishing 28.8 million exchanges (prompts and completions) with Claude.
- The Targeted Capabilities: Rather than simple conversational data, the campaign specifically targeted Anthropic’s most complex, frontier-class capabilities. These included agentic reasoning, advanced software engineering, and long-horizon tasks—benchmarks heavily associated with Anthropic’s advanced Mythos architecture.
[25,000 Fake Proxy Accounts] ──► [28.8M Claude Exchanges YTD] ──► [Targeted Harvesting of Agentic & Coding Tech] ──► [Repackaged into Alibaba Qwen Models]
2. The Mechanics of “Adversarial Distillation”
At the core of the legal and political dispute is a practice known in machine learning as model distillation.
While distillation is a standard, often legal practice used globally to train smaller, cheaper models by learning from the outputs of a larger AI, Anthropic stresses that executing it at an industrial scale via fraudulent accounts violates its Terms of Service.
Free-Riding on R&D: In its letter, Anthropic warned that Alibaba is engaged in unauthorized harvesting to build rival chatbots at a mere fraction of the cost, stating: “These attacks harvest U.S. AI capabilities across frontier labs and repackage them as their own without incurring the training and R&D costs required to train U.S. frontier models.”
3. Part of a Growing Trend
The allegation marks the second major wave of distillation warnings sounded by Anthropic this year. In February, the lab publicly flagged three other Chinese AI entities for using similar tactics, though Alibaba’s alleged campaign dwarfs them all combined:
| Chinese AI Entity / Laboratory | Estimated Claude Interactions (Reported) | Fake Proxy Accounts Utilized |
| Alibaba (Qwen Lab) | 28.8 Million (April–June 2026) | ~25,000 |
| MiniMax | 13.0 Million (Pre-February 2026) | ——— |
| Moonshot AI | 3.4 Million (Pre-February 2026) | ——— |
| DeepSeek | 150,000+ (Pre-February 2026) | Shared pool of ~24,000 accounts across early labs |
4. Intense Washington Political Blowback
The timing of the letter could not be more delicate for Alibaba, which just days ago filed a federal lawsuit against the U.S. Pentagon to challenge its inclusion on the Section 1260H military-linked blacklist.
Anthropic’s public push is already driving policy reactions in Washington. Top U.S. labs—including Anthropic, OpenAI, and Google—have quietly formed an alliance to share real-time threat intelligence regarding cross-border distillation attempts.
Furthermore, Anthropic is explicitly urging the U.S. government to step up restrictions, close loopholes that allow Chinese entities to rent cloud-based U.S. computing power, and implement strict amendments to penalize or blacklist any foreign laboratory found to be structurally “stealing” American model outputs.
