共有:

Claude · AI IP Dispute

Anthropic Names Alibaba in 28.8M Distillation Attack

Alibaba-linked accounts allegedly queried the Claude API 28.8 million times to harvest training data for competing models — in direct violation of Anthropic's Terms of Service. The US-China AI race has entered IP lawsuit territory.

AI Navigate Editorial·2026.06.26·6 min read

Claude API 28.8M queries output harvest Alibaba-linked accounts rival model training data ToS violation: "no training rival models on Claude output"
01
What Is Distillation

What a Distillation Attack Is

Model distillation is the technique of collecting large volumes of output from a capable "teacher model" and using it to train a smaller "student model" — transferring capabilities without access to the original weights. It's a legitimate research technique, but harvesting another company's proprietary API output without consent to train a competing model is explicitly banned in nearly every major AI provider's Terms of Service.

Alibaba-linked accounts allegedly queried the Claude API more than 28.8 million times to collect outputs as training material for a competing model, according to Anthropic. This is the first high-profile case in which Anthropic has publicly named a specific company in a ToS violation of this type.

02
The Numbers

Scale of the Alleged Attack

28.8M+
API queries alleged
First
named-company public accusation
ToS §1
rival training prohibition
03
Enterprise Checklist

What Enterprises Should Do Now

01

Re-read your AI API Terms of Service

OpenAI, Anthropic, Google, and every major AI provider explicitly prohibits using API output to train competing models. If your team is fine-tuning internal models, verify the provenance of your training data.

02

Assess procurement risk for China-linked AI

As US-China IP disputes escalate, using AI APIs from China-linked vendors may carry legal risk in regulated industries or US government procurement contexts. Flag this for legal review early.

03

Monitor for anomalous API call volumes

Check that outsourced vendors or internal automation scripts aren't generating high-volume API calls. You don't want to be an inadvertent participant in a ToS violation at scale.