Tensions are escalating between US-based Anthropic and China's Alibaba in the latest chapter of fierce AI competition, after reports that the Chinese company decided to ban its employees from using the programming tool Claude Code, just days after Anthropic accused entities linked to Alibaba of trying to extract secrets from the Claude model through millions of queries.

Anthropic accuses Alibaba

Anthropic recently accused entities linked to China's Alibaba and its AI company Qwen of conducting a broad campaign to extract capabilities from the Claude model by directing millions of queries at it, marking a new confrontation among major AI companies over protecting advanced models and intellectual property rights.

According to a Reuters report, Anthropic claimed in a letter to US lawmakers that the entities used around 25,000 fraudulent accounts to conduct over 28.8 million interactions with Claude, aiming to gather detailed information about its workings and internal capabilities.

Alibaba has not issued any official comment on these allegations as of yet, nor has any independent source confirmed the claims.

The technique of 'model distillation'

Anthropic's allegations are based on a technique known as 'model distillation,' a method AI companies typically use to create smaller, more efficient models based on their original models.

Anthropic asserts that using a competitor's model for this purpose without permission constitutes an intellectual property violation, noting that the alleged accounts posed complex and intensive questions about Claude's capabilities in software engineering and advanced reasoning, allowing them to form a precise picture of its behavior and workings without accessing its source code or training data.

The case highlights a growing challenge for AI model developers, as much of any advanced model's behavior can theoretically be reconstructed by conducting a vast number of carefully crafted interactions with it, potentially reducing the need for billion-dollar investments to develop competing models from scratch.

This is not the first time Anthropic has raised this issue; earlier this year, it made similar accusations against other Chinese companies, notably DeepSeek, as well as Moonshot AI and MiniMax, while OpenAI has also expressed similar concerns about its models being exploited in this way.

Anthropic believes that continuation of these practices could weaken the incentive for innovation, if leading models can be imitated by analyzing their responses instead of developing new technologies.

The company urged US lawmakers to take swift action to curb these practices, arguing that future competition in AI may not be limited to developing the most intelligent models, but will extend to protecting models from attempts to analyze and copy their behavior through millions of questions.

Secret code in Claude... and Alibaba bans it

According to media reports, China's Alibaba plans to ban its employees from using Anthropic's programming tool Claude Code starting July 10.

This step comes amid ongoing restrictions imposed by Anthropic on making its models available to Chinese companies, as the company has been working for months to close loopholes that allow users in China to access Claude model services.

Media reports revealed that one of the protective measures tested by Anthropic included a version of Claude Code capable of recognizing users in China through a secret mechanism.

Thariq Shihab, a company official, explained in a post on platform X that this mechanism was part of an experiment launched last March to limit misuse of unauthorized accounts, as well as to protect models from 'model distillation' operations. Shihab added that the company later implemented stronger protective measures and had been planning to remove that experimental mechanism for some time.

According to reports, Alibaba classified the Claude Code tool as high-risk software and directed its employees to use the internally developed Qoder tool instead.

Tags

Anthropic Artificial Intelligence Alibaba

Link copied

July 8, 2026 | Last updated: July 8, 2026

2 minutes

Follow us