Every student knows the move: glance over a smarter classmate's shoulder, copy the answers, and hope the teacher doesn't notice the suspiciously identical wrong answers. Anthropic now says one of the biggest names in tech tried exactly that — at a scale that makes the average cheat sheet look quaint.

29 Million Suspicious Conversations

Anthropic has publicly accused Alibaba of running the largest known "distillation" campaign against its Claude models. According to Anthropic, the effort used thousands of fraudulent accounts spread across nearly 29 million interactions, all aimed at extracting Claude's capabilities in software engineering and agentic reasoning.

Distillation, in plain terms, is using one model's outputs to train another — effectively teaching a student model to imitate a teacher it never paid for. Do it at the scale Anthropic is alleging, and you're not sampling for research. You're trying to clone the thing.

Distillation, or Theft With Extra Steps?

This is the frontier labs' dirty little open secret finally getting aired in public. Training a model from scratch costs a fortune in compute and talent; quietly siphoning a rival's answers is dramatically cheaper, which is precisely why everyone worries about it and few say it out loud. Anthropic naming Alibaba directly is the unusual part.

It also exposes how flimsy the moats really are. If a frontier model's hard-won abilities can be approximated by funneling enough cleverly disguised queries through it, then the lead any lab holds is measured in months, not years — and the security teams guarding those APIs just became as important as the researchers building the models.

The AI arms race has officially entered its industrial-espionage era. Lawyer up, everyone.

Source: Tech Startups