DeepSeek and Moonshot turn to OpenAI’s GPT-6 Astra for training data

How distillation works

Distillation can give a lab training data by eliciting a model’s hidden reasoning steps. OpenAI and Anthropic prohibit the practice.

Moonshot declined to comment to The Information. DeepSeek did not respond to the publication’s request.

A shifting target

OpenAI said in late September that it had identified a coordinated distillation campaign involving people it said were connected to Moonshot. A company representative said OpenAI takes the issue very seriously.

But the company has also said attackers changed tactics each time it introduced a new safeguard. That makes Astra’s reported use more than a question of whether one model can be protected: defenses may push extraction attempts toward another system rather than stop them.

I think the important omission is how OpenAI’s latest protections perform against the tactics now being used. The report describes the shift from Claude to Astra, but offers no measure of how much Astra data the labs obtained—or whether OpenAI’s defenses reduced that amount. Until those numbers are clear, a change of target is evidence of adaptation, not evidence that safeguards have failed.