Anthropic Alibaba distillation accusation
“Carried out illicitly, systematically, and at an industrial scale.” That is how Anthropic described what it says happened to its own Claude models in a letter sent to the US Senate Banking Committee on June 10. The company accused operators affiliated with Alibaba and its Qwen AI lab of running the largest model-extraction campaign it has ever documented against itself, nearly 25,000 fraudulent accounts, 28.8 million exchanges with Claude, over six weeks this spring.
Two days after that letter went out, the same US government imposed export restrictions on Anthropic’s own most advanced models, the story we have been following closely since it broke.
| VERDICT: A genuinely large, well-documented accusation, sitting inside a deeply ironic timeline Anthropic itself has to navigate carefully. Anthropic alleges the campaign, run between April 22 and June 5, specifically targeted Claude’s software engineering and agentic reasoning abilities, its most commercially valuable skills, using a technique it calls adversarial distillation. This dwarfs three earlier extraction campaigns Anthropic disclosed in February, attributed to DeepSeek, Moonshot AI, and MiniMax. The genuinely sharp detail: Anthropic sent this letter asking Washington to help it fight Chinese extraction of its technology two days before that same Washington restricted foreign access to that technology itself. |
What Anthropic Is Actually Alleging
The letter, reviewed by Reuters, Bloomberg, and CNBC, states that operators connected to Alibaba’s Qwen lab deployed thousands of fraudulent user accounts specifically to evade Anthropic’s usage restrictions and access Claude at scale. The technique described, adversarial distillation, works by repeatedly prompting an advanced model to extract its reasoning patterns and underlying behavior, then using those responses to train a competing model without bearing the original research and development cost.
Anthropic’s letter draws a direct line to safety, not just commercial harm: when a frontier model is cloned through unauthorised distillation this way, the resulting copy typically strips away the safety guardrails engineered into the original model, since those protections are not what the extraction process is designed to capture. The company argues this is precisely why distillation deserves treatment as a security issue, not simply an intellectual property dispute between competitors.
Scale is the detail that separates this from prior cases. Anthropic’s own February disclosure named DeepSeek at roughly 150,000 exchanges, Moonshot AI at 3.4 million, and MiniMax at more than 13 million. The alleged Alibaba campaign, at 28.8 million exchanges through 25,000 fake accounts, is described in Anthropic’s own letter as the largest known distillation attack on the company to date, by a wide margin over anything documented before it.
The Timing Anthropic Now Has to Live With
This is the part of the story that deserves more attention than the raw numbers alone. Anthropic’s letter, dated June 10, explicitly asked US lawmakers to strengthen enforcement against foreign extraction of American AI capabilities, framing protection of US models and their commercial deployment as complementary goals, not competing ones.
Two days later, on June 12, Commerce Secretary Howard Lutnick signed the order blocking foreign nationals from accessing Anthropic’s own latest models, Fable 5 and Mythos 5, the export control order we covered in detail at the time, and the same order President Trump later softened his public tone on after meeting Dario Amodei at the G7. Anthropic is now in the position of needing the same government it is currently fighting an access restriction with to simultaneously crack down harder on a different country extracting its technology. The company’s own framing, that defending against distillation and allowing commercial deployment are complementary, is the argument it needs Washington to accept on both fronts at once, and the government’s actions in the same two-week window suggest that argument has not yet fully landed.
Why This Connects to Everything Else in This Story
This is now the fourth thread in a single, continuous month for Anthropic specifically. The NSA reportedly breached almost all of its own classified systems testing Anthropic’s Mythos model, in hours rather than weeks. The Commerce Department restricted foreign access to that same model family days later, reportedly triggered by a separate jailbreak Amazon discovered, with Amazon itself being both an $8 billion investor in Anthropic and a commercial rival running competing models. Trump softened his public stance after meeting Amodei at the G7, while the underlying restriction stayed fully in force. And now Anthropic is publicly accusing a major Chinese tech conglomerate of the largest extraction campaign it has ever recorded, two days before its own products got cut off by the government it was asking for help.
Read together, the throughline is the same one we have been tracking since the G7 lunch: frontier AI access is being negotiated, restricted, and fought over simultaneously, by multiple parties, often working at cross purposes even within the same government, in roughly the same six-week window. Anthropic is not a passive bystander in this story. It is simultaneously the company asking for tougher export enforcement against China and the company whose own export access just got tightened by the same enforcement apparatus.
Why This Matters for the UAE Specifically
We have made the case throughout this series that the UAE’s AI strategy depends on stable, continued access to the same US frontier model ecosystem now sitting at the centre of this dispute. This accusation adds another layer to that picture. Washington’s tightening posture toward Chinese access to American AI capability, illustrated here by Anthropic’s own complaint, is the same broader instinct shaping the export controls that already restricted UAE access to Mythos and Fable, the structured-access coalition Amodei and Hassabis pitched at the G7, and the trusted-partner tiering question we have flagged repeatedly. Every escalation in the US-China AI access fight tightens the overall framework the UAE’s own access sits inside, regardless of which specific country is named in any individual dispute.
Alibaba’s Side
Alibaba had not responded to requests for comment from any of the outlets that reviewed the letter at the time of their reporting. Separately and worth noting for context, Alibaba was added to the Pentagon’s list of Chinese military companies this month, a designation the company is reportedly challenging, adding another live dispute to the company’s relationship with US regulators alongside this specific accusation.
Sources
• CNBC: Anthropic accuses Alibaba of campaign to ‘brazenly’ and ‘illicitly’ extract AI capabilities — https://www.cnbc.com/2026/06/24/anthropic-alibaba-distillation-campaign.html
• Benzinga: Anthropic accuses Alibaba of AI model theft, says 28.8 million Claude conversations were harvested — https://www.benzinga.com/markets/tech/26/06/60088712/anthropic-accuses-alibaba-of-ai-model-theft-says-28-8-million-claude-conversations-were-harvested
• Business Today: Anthropic accuses Alibaba of largest ever attempt to extract Claude AI capabilities — https://www.businesstoday.in/technology/artificial-intelligence/story/anthropic-accuses-alibaba-of-largest-ever-attempt-to-extract-claude-ai-capabilities-539087-2026-06-25
• Edgen: Anthropic accuses Alibaba of $29M illicit AI model extractions — https://www.edgen.tech/news/post/anthropic-accuses-alibaba-of-29m-illicit-ai-model-extractions
Robius.news — Dubai, UAE — 2026 | Built to be first. Built to be trusted.
