Is imitation the sincerest form of flattery?

Have you ever been tempted to copy someone elses homework, rather than do the work yourself?

Well Anthropic (Claude) found that three Chinese models have been trying to do just that, and reverse engineer Claude through how it reasons and solves problems.

Using 24,000 fake accounts, DeepSeek, Moonshot, and MiniMax ran coordinated 16M+ exchanges with Claude to understand how Claude does what it does.

They accused them of distillation, which involves training a weaker system on the outputs from a stronger one.  One of their biggest concerns is a moral and ethical one - that the Chinese models will not contain the safeguards that Anthropic have in place.

MiniMax ran the largest campaign at 13M+ exchanges, fortunately Anthropic caught it mid-operation, and issued a new release in under 24 hours.

Meanwhile DeepSeek had Claude spell out reasoning step-by-step and rewrite politically sensitive queries, generating training data for both logic and censorship.

Open AI raised similar concerns with US Congress weeks ago.

The Chinese are known as fast followers, and this is apparent in how they are using the learnings of others to accelerate their own capabilities.Ironically, many of the frontier models (Chat GPT, Claude) were trained on IP and data that belonged to others, and have since paid settlements and licenses for their use.

SOURCE

Article: https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks