Anthropic, the AI company behind the Claude language model, has come forward with serious accusations. It claims that three Chinese AI developers—DeepSeek, Moonshot, and MiniMax—have been running large-scale campaigns to steal capabilities from Claude. These campaigns involved using deceptive tactics to extract knowledge from the model without permission. The incident raises questions about how AI models are built and protected.

How the campaigns unfolded

Anthropic states that the three companies used a method called distillation. This process involves training a less advanced model on the outputs generated by a more powerful one. By doing this at scale, the companies aimed to replicate Claude’s reasoning, tool use, and coding skills. They set up around 24,000 fake accounts to interact with Claude, generating over 16 million exchanges in total.

The companies reportedly used proxy services to access Claude, which they did without authorization. These services resell access to AI models at large scale, often through complex architectures called hydra clusters. The campaigns were designed to evade detection, with patterns of activity that looked coordinated. For example, accounts shared payment methods and timed their interactions to mimic load balancing and avoid raising suspicion.

Details of the individual campaigns

The DeepSeek campaign involved more than 150,000 exchanges. Its main goal was to extract Claude’s reasoning skills across various tasks. The activity showed synchronized traffic, with similar patterns across multiple accounts. Meanwhile, Moonshot AI conducted over 3.4 million exchanges focused on learning how Claude reasoned about tools, coding, and data analysis. They also aimed to reconstruct how Claude arrives at its answers.

The largest of the three, MiniMax, carried out over 13 million exchanges. Its primary focus was on Claude’s ability to code and orchestrate tool use. Anthropic detected MiniMax while it was active. Remarkably, nearly half of its traffic was redirected to Claude’s new model within just 24 hours of detection, indicating a rapid adaptation by the attackers.

Broader questions about AI training and security

Anthropic claims these campaigns relied heavily on commercial proxy services that resell access to models like Claude. This highlights a larger issue in AI development: how models are trained and protected. Many large language models are trained on vast amounts of internet data, often without explicit consent from the original content creators. This practice raises concerns about data ownership and ethics.

Experts note that this situation underscores ongoing debates about the legality and morality of using publicly available data for training AI systems. Since models are often built by indexing the internet’s content, there is little clarity on who owns the resulting synthetic data. The allegations against the Chinese firms bring to light the risks of unauthorized data extraction and model theft in the AI industry.

Overall, the incident shines a light on the vulnerabilities of large language models. As AI models become more powerful and accessible, protecting them from illicit copying and misuse becomes increasingly important. This case may lead to new discussions on how to secure AI systems and enforce better regulations to prevent similar activities in the future.

Inspired by

• https://www.computerworld.com/article/4136474/anthropic-alleges-large-scale-distillation-campaigns-targeting-claude-2.html

Sources

• anthropic.com