American AI research company Anthropic has accused its Chinese counterpart Alibaba of illegally exploiting and accessing its popular AI model, Cloud.

According to the Financial Times, the American company sent an official indictment letter to the US Congress, describing what happened as the largest "distillation and inference" attack on its services and technology platforms.

Millions of interactions and fake accounts

The American company explained in its indictments that Alibaba used and activated more than 25,000 fake accounts, to generate approximately 28.8 million messages or direct interactions with its artificial intelligence tools. This systematic step was a direct and explicit breach of Anthropic's Terms of Service and Use.

Technical indicators indicated that this targeted attack lasted between April 22 and June 5, and was primarily aimed at gaining access to the advanced capabilities of the company's Mythos model, before the US government issued a decision to block access to this sensitive model.

Anthropic has made a direct and detailed accusation against Alibaba and its artificial intelligence lab known as Alibaba Qwen.

The company called on the US Senate to take the necessary legislative and security measures and measures to prevent such breaches, and to punish the parties and companies responsible for them. In its letter, it also expressed its full support for the U.S. government's efforts to combat digital theft attacks, including building strong partnerships with private sector AI companies to enhance national technology security.

Anthropic has previously faced similar attacks from multiple Chinese companies that used distillation and inference techniques to train and develop their AI models by hacking into Cloud tools.

The report released at the time included monitoring the volume of operations, as the "DeepSec" campaign included more than 150,000 exchanges, while "Moonshot AI" recorded about 3.4 million exchanges, and "MiniMax" broke the barrier of 13 million illegal exchanges for stealing training data.

Alibaba Group has been placed on the list of companies banned by the US government on suspicion of being linked to the Chinese Liberation Army and providing support for military activities against US targets. Alibaba has denied any links to the Chinese military, but has remained silent and has not responded to Reuters' request for comment on the recent technical accusations made against it by Anthropic.

What are Drifting Attacks?

A report by Forbes magazine described distillation attacks as the latest and most dangerous weapon in the ongoing artificial intelligence war between the United States and China, and its danger lies in the fact that it allows attackers to exploit the sophisticated cognitive capabilities of any leading AI model in order to train and develop another replica model at a low cost.

The technical and economic mechanism of these attacks can be summarized in the following points:

Data collection mechanism: The attacking company creates accounts and generates millions of questions to get typical answers from the AI (the victim).

Feeding the simulation model: Emerging AI models are fed these integrated pairs (advanced question and answer) and used as a ready-made database for training.

Provisioning compute and processing: This convoluted method saves businesses from the need to import and purchase large volumes of raw data, and reduces the superior computing power and complex processing required for initial setups.

Economic outcome: Launch of powerful AI models with the same capabilities as the original models, but at a much lower production and operational cost as a result of avoiding the high incorporation cost