Latest / Minds, Bodies, and Terawatts / Anthropic Accuses Alibaba of Largest-Ever AI Distillation Attack Using 25,000 Fake Accounts
Transcript
- 0:00[Host] This is Minds, Bodies, and Terawatts, where the AI arms race goes undercover. CNBC confirmed on June 24th that Anthropic has accused Alibaba of running the largest known distillation attack against its Claude AI, using 25,000 fake accounts to generate nearly 29 million conversations.
- 0:18[Guest] Anthropic detailed the campaign in a letter to Senators Tim Scott and Elizabeth Warren, sent June 10th ahead of a Senate hearing on AI and the American Dream. The operation ran from April 22nd to June 5th, with operators linked to Alibaba's Qwen AI lab targeting Claude's most advanced capabilities.
- 0:38[Host] That is an industrial-scale operation. What specifically were they after?
- 0:42[Guest] Mythos Preview, Anthropic's frontier model for agentic reasoning, software engineering, and long-horizon tasks. Distillation feeds a weaker model the outputs of a stronger one until the student mimics the teacher, and Alibaba allegedly ran it at massive scale to boost Qwen.
- 0:59[Host] This was not the first time.
- 1:01[Guest] In April, Anthropic flagged DeepSeek, Moonshot, and MiniMax for a similar campaign with 24,000 accounts and 16 million exchanges. The White House called it industrial-scale AI theft. The Alibaba operation nearly doubled that volume.
- 1:15[Host] OpenAI and Google have flagged similar operations against their own models.
- 1:20[Guest] Which makes this an industry-wide vulnerability. As long as a model has a public API, anyone with a credit card can extract its reasoning patterns at a fraction of the training cost.
- 1:33[Host] What is Anthropic asking Congress to do?
- 1:36[Guest] Their letter described distillation as a way to harvest U.S. AI capabilities across frontier labs and repackage them without incurring the training costs. That framing gives lawmakers a trade-policy hook to restrict who accesses frontier models.
- 1:51[Host] But two days after Anthropic sent that letter, the Commerce Department imposed export controls on Anthropic's own Mythos and Fable models. The company asking Congress to protect its AI had to disable global access to that same AI.
- 2:06[Guest] Which exposes the deeper bind. Restricting access protects the model from distillation but shrinks the customer base funding the next generation. The labs need worldwide revenue to stay ahead, and Washington is narrowing who can pay.
- 2:21[Host] Meanwhile the investment at risk keeps growing. xAI's Memphis campus runs over 400,000 GPUs, and Meta guided $135 billion in capital spending this year.
- 2:31[Guest] All to train models that fake accounts were querying for pennies. The gap between what it costs to build a frontier model and what it costs to copy one is the vulnerability no firewall can patch.
- 2:45[Host] And defenses like account verification and rate limiting are the only technical barrier?
- 2:51[Guest] The fake accounts blew right past them. The better the model gets, the more valuable distillation becomes, because at the query level the traffic looks like normal usage.
- 3:02[Host] So dominance and vulnerability grow in lockstep.
- 3:05[Guest] The same concentration that makes American labs the most powerful AI producers on Earth makes them the highest-value targets for extraction.
- 3:14[Host] Patrick Deglon describes this dynamic as the Star Wars path in his book Unscarcity, where superpowers compete for AI dominance and the tools themselves become the battlefield.
- 3:25[Guest] And when a credit card and a script can harvest what hundreds of billions built, the race is no longer about who builds the best model. It is about who keeps it. For more on how AI concentration reshapes global power, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.