- Anthropic accused Alibaba’s Qwen AI lab of conducting 28.8 million exchanges with Claude through nearly 25,000 fraudulent accounts between April and June — calling it the largest attempt by a Chinese company to “illicitly” access a US frontier AI model via adversarial distillation.
- The campaign specifically targeted Claude’s highest-value capabilities including software engineering and agentic reasoning, occurring after the White House OSTP issued a memo in April warning against such practices — making it a direct act of defiance against stated US policy.
- Alibaba shares fell as much as 4.8% in Hong Kong following the news; the company had no comment and is already on the Pentagon’s military-company blacklist, which Anthropic cited in its letter.
- Senate legislation (Hagerty/Kim) is being drafted to blacklist or sanction Chinese firms using distillation to steal US AI model output, and a related House bill (Huizenga/Kamlager-Dove) is also being considered for inclusion in the annual defense bill.
What Happened?
Anthropic sent a letter to several US senators and White House officials accusing Chinese tech giant Alibaba of orchestrating a large-scale “adversarial distillation” attack against its Claude AI model. According to people familiar with the document and a copy seen by Bloomberg, the campaign involved 28.8 million exchanges with Claude conducted through nearly 25,000 fraudulent accounts between April and June — the most extensive such attempt Anthropic has documented from a Chinese company. The attacks specifically targeted software engineering and agentic reasoning — Claude’s most commercially valuable capabilities. Anthropic said it was the biggest attempt by a Chinese company to piggyback on US AI work, and it follows previously disclosed distillation efforts by DeepSeek and Minimax.
Why It Matters?
Adversarial distillation allows a rival to replicate the capabilities of a frontier AI model at a tiny fraction of the training cost by flooding the target model with carefully crafted queries and using the outputs to train a new model. If done at scale, it effectively transfers years of R&D investment to a competitor without authorization. US officials have estimated unauthorized distillation costs Silicon Valley labs billions annually. For Anthropic specifically — valued at $965 billion by private investors and preparing for an IPO — this is an existential competitive threat: cheaper Chinese models trained on Claude’s outputs could undercut its pricing and erode the moat its frontier capabilities are supposed to provide. The political backdrop is complex: Anthropic is simultaneously seeking government help against distillation while embroiled in a dispute with the Trump administration over export controls on its own top-tier Fable 5 and Mythos 5 models.
What’s Next?
Senate legislation is being fast-tracked: Sens. Hagerty (R-TN) and Kim (D-NJ) plan to introduce an amendment to the defense bill as soon as this week that would blacklist or sanction Chinese firms caught using distillation to copy US models. A parallel House bill is also in play. Anthropic is asking the administration to clarify antitrust guidelines to allow US labs to share more information about distillation attacks, tighten AI chip export controls, and impose penalties on distillation violators. The irony is that Anthropic needs Trump administration cooperation on all of this while simultaneously fighting the administration’s restrictions on its own advanced models — making the two disputes hard to decouple diplomatically.
Source: Bloomberg
