Anthropic has accused three Chinese AI companies of running massive campaigns to illegally extract capabilities from its Claude AI model. The company claims these campaigns involved millions of interactions and used deceptive tactics to bypass restrictions. This controversy highlights ongoing issues around AI model security and data use.

How the Campaigns Worked

Anthropic says the three companies—DeepSeek, Moonshot, and MiniMax—used a method called distillation. This process involves training a simpler model on outputs generated by a more advanced AI. In this case, they used false accounts and proxy services to access Claude at scale, evading detection and regional restrictions.

DeepSeek’s effort involved over 150,000 exchanges, focusing on extracting reasoning skills across different tasks. The activity showed patterns of synchronized traffic, with accounts sharing payment methods and timing, likely to spread out their access and avoid suspicion. Moonshot undertook over 3.4 million exchanges, targeting reasoning, tool use, coding, and data analysis. MiniMax was the largest, with more than 13 million exchanges, mainly aimed at coding, tool use, and orchestrating AI functions. Notably, MiniMax redirected nearly half of its traffic to a newer Claude model within a day of its release.

Methods and Tools Used in the Theft

The companies relied heavily on commercial proxy services, which resell access to Claude and other advanced AI models at scale. These services used complex architectures called hydra clusters, making it easier to mask the true origin of their requests and handle high volumes of data. This approach allowed them to carry out large-scale extraction without immediate detection.

Anthropic claims that these tactics violate their terms of service and regional access restrictions. Despite these efforts, the company managed to detect and disrupt the campaigns while they were ongoing. The use of proxy services and coordinated account activity highlights the challenges AI companies face in protecting their models from misuse and theft.

Broader Issues in AI Development

These allegations raise bigger questions about how AI models are trained and the legalities involved. Most large language models are trained on vast amounts of internet data, often without explicit permission from content creators. This practice is common but controversial, especially when models are distilled or optimized based on such data.

Experts note that the line between legal data use and unauthorized extraction is blurry. Neil Shah from Counterpoint Research points out that many models are built by indexing the internet, sometimes using data without clear consent. The debate over who owns synthetic data used for training and whether it’s acceptable to use it remains unresolved and controversial.

This case underscores the ongoing struggle to balance innovation, security, and legal boundaries in AI development. As models become more advanced, so do the methods to steal or misuse them, prompting calls for tighter controls and better safeguards in the industry.

Inspired by

• https://www.infoworld.com/article/4136468/anthropic-alleges-large-scale-distillation-campaigns-targeting-claude.html

Sources

• anthropic.com