Anthropic accuses Alibaba of 'illicitly' accessing Claude models

Anthropic, per a Bloomberg report, has formally complained to US senators and White House officials that operators it links to Alibaba's Qwen AI lab 'illicitly' accessed Claude to extract its most valuable capabilities, specifically software-engineering performance and agentic reasoning. Anthropic framed it as the most aggressive attempt yet by a Chinese firm to 'piggyback' on the work of leading American labs — language designed to resonate in a Washington already primed for export controls and national-security scrutiny of frontier AI.

The core technical concern is distillation: using API access to a frontier model to generate training data that transfers its behavior into a cheaper, locally-controlled model. Because the alleged extraction targets agentic and coding skills — the exact areas where Claude leads — the claim spotlights a structural vulnerability facing every API-access vendor. If a competitor can systematically probe and replicate a model's edge through ordinary queries, the moat erodes regardless of weights staying private.

The accusation arrives in a charged context. Alibaba just debuted its Qwen Robot Suite of embodied models, and Chinese open-weight efforts like DeepSeek and GLM are drawing praise from Western developers. It also dovetails with Anthropic's commercial moment: the company launched Claude Tag this week and is reportedly carrying a ~$965B valuation ahead of a confidential IPO filing, giving it both motive and platform to press regulators.

Skeptics will note the claim is, so far, an assertion to officials rather than a filed lawsuit with public evidence, and that 'distillation' accusations are notoriously hard to prove. Watch for whether the US acts — and whether Alibaba responds — as this becomes a test case for how IP norms apply to model behavior rather than code or weights.