Alibaba Accused of Running 25,000 Fake Accounts to Steal Claude's Capabilities via Distillation

[Warden]

Anthropic told US senators this week that Alibaba-linked operators ran nearly 25,000 fake accounts to copy Claude's capabilities through what the company calls illicit distillation. The alleged campaign logged 28.8 million exchanges between April 22 and June 5 2026, targeting Claude's Mythos Preview model. Anthropic tied the effort to previous campaigns it has blamed on DeepSeek, Moonshot AI and MiniMax, all of which deny the allegations.

Model distillation at this scale is a significant concern for the AI industry. The basic technique involves running huge numbers of inference queries against a target model, using the outputs as training data to train a new model that mimics the target model's behaviour. Done systematically enough it allows a competitor to effectively approximate a frontier model's capabilities without doing the underlying training work themselves. The 28.8 million exchange number suggests this was highly organised rather than opportunistic.

The White House Office of Science and Technology Policy had already accused foreign entities of industrial-scale campaigns of this type back in April. OpenAI separately warned lawmakers about similar distillation attempts against its models. The fact that multiple major US AI companies are raising these concerns simultaneously with Senate testimony suggests this is being treated as a national security issue as much as a commercial one, and it is worth noting this is part of the background context for the Fable 5 export ban.

Anthropic accuses Alibaba of campaign to 'brazenly' and 'illicitly' extract AI capabilitieswww.cnbc.com

[ProperJobs50]

28.8 million exchanges in about six weeks is industrial scale. That is not a research team exploring a competitor's model, that is an organised programme with engineering resources behind it

[StoneCold_99]

The legal situation here is murky. Terms of service violations are clear but whether this constitutes IP theft in a legally actionable sense is genuinely unclear and courts have not caught up

[HeartbreakKid_Fan]

All the named companies deny it. But what else would they say? The practical question is whether Anthropic can prove attribution convincingly enough to matter in a US Senate hearing

[Holly43]

Distillation at this scale is the reason the export ban makes some sense even if the implementation was clumsy. If Mythos capabilities can be extracted this thoroughly, keeping the model behind nationality checks matters

[Tara_66]

The fact that DeepSeek, Moonshot and MiniMax are all named from previous incidents suggests Anthropic has been tracking this for a while and the Alibaba accusation is not a first offence claim

[IronQuarry48]

I am sceptical that distillation alone can reproduce the full capabilities of a frontier model. You can get a lot of surface behaviour but the deeper reasoning patterns and knowledge are harder to extract this way

[Josh_79]

If Anthropic has attribution evidence strong enough to share with Senate committees, why have there been no legal actions against the named companies? The jurisdictional complexity is the obvious answer

[Donna]

25,000 fake accounts is a significant operational security failure on Anthropic's side as much as anything else. How were these accounts not detected and closed down during the six-week campaign

[CMPunk_Mike]

The OSTP framing of this as industrial scale in April and now Anthropic naming specific actors in June suggests we are moving toward some kind of formal legal or diplomatic response