Capability Extraction at Scale: Anthropic Says Frontier Models Were Targeted for Distillation

The frontier AI race is entering a new phase. The competition is no longer only about who can build the most powerful model, but who can learn the most from the models that already exist.

Anthropic says operators affiliated with Alibaba’s Qwen division, DeepSeek, Moonshot AI and other organizations carried out one of the largest distillation campaigns it has detected to date. Between April 22 and June 5, 2026, nearly 25,000 fraudulent accounts generated more than 28.8 million interactions with Claude models, targeting capabilities in coding, autonomous agents, complex planning and advanced reasoning.

The company says the goal was distillation: using the outputs of a frontier model as synthetic training data to help develop competing systems at significantly lower cost. Anthropic detailed the findings in a letter sent to U.S. senators and White House officials, arguing that the incident highlights a growing challenge for frontier AI companies: protecting valuable capabilities in systems that become more powerful every time they are used.

The New AI Shortcut

Building a frontier model requires enormous investment. Companies need advanced chips, massive data centers, specialized researchers and billions of dollars in training costs. The largest AI systems are the result of years of accumulated infrastructure and experimentation.

Distillation creates a different path. Instead of rebuilding a model from the beginning, developers can study the behavior of an existing system and use its outputs to train another model.

The process does not copy the original model weights. Instead, it attempts to transfer useful capabilities, such as reasoning patterns, coding approaches or problem-solving strategies, into a new system.

That makes distillation attractive. It turns a finished model into a source of training data.

For frontier AI companies, that creates a difficult problem: the more useful a model becomes, the more information it reveals through interaction.

The Weakness of the AI Moat

Traditional software companies protect their technology by protecting their code. AI companies face a different challenge.

A model’s value is not only inside its architecture. It appears in the answers it produces. Every generated response can reveal something about how the system reasons, writes code or approaches complex tasks.

That creates a new kind of competitive risk. A company can spend billions developing a frontier model, only to find that competitors may be able to learn from the system without reproducing the original training process.

The moat is no longer just the model itself. It is also the ability to control access to what the model can do.

A New Layer in the AI Race

Anthropic has previously disclosed other distillation-related campaigns involving companies including DeepSeek, Moonshot AI and MiniMax. The latest disclosure expands the scale of the issue and places AI capability extraction into a wider debate about competition, security and regulation.

The concern extends beyond one company or one country. As frontier models become more capable, access to those systems becomes strategically valuable.

Governments are already competing over chips, infrastructure and AI leadership. Companies are now confronting another question:

How do you protect intelligence once it can be observed?

The Cost of Open Intelligence

AI systems are different from traditional technology because interaction is part of the product. A search engine can return a result. A database can provide information.

A reasoning model demonstrates a process. That process is exactly what makes advanced AI systems valuable and what makes them difficult to contain.

The industry is moving toward a world where access itself becomes a strategic decision. Companies must balance usefulness with control. More users create more value, but more exposure also creates more opportunities for imitation.

The Grey Terminal Note

The AI industry has spent years treating model size, compute and data as the foundation of competitive advantage. But frontier intelligence introduces a different problem.

The most valuable asset may not only be the model sitting inside a data center. It may be the capability revealed every time someone interacts with it.

The next AI race will not only be fought over who can build the smartest system. It will also be fought over who can protect the intelligence their systems produce.