Chinese AI developers have shifted their training focus to OpenAI’s GPT-6 Astra model following stricter security measures from Anthropic. Reports indicate that DeepSeek and Moonshot AI relied heavily on Astra’s outputs to train their latest systems, including Kimi K3.1 Pro and V4.1-Flash 1017. This approach requires extracting hidden reasoning steps from the target model, a process both OpenAI and Anthropic explicitly forbid. While Anthropic claims its new methods reduce the data attackers can extract by up to 99 percent, OpenAI noted a coordinated distillation campaign involving Moonshot-linked individuals in late September. Previously, Claude served as the primary target for these operations, but the company has since tightened its defences significantly.
The shift highlights the ongoing cat-and-mouse game between model providers and researchers attempting to replicate proprietary capabilities. OpenAI stated it addresses the issue seriously, though past adjustments to defences often led attackers to adapt their tactics quickly. Neither Moonshot nor DeepSeek responded to requests for comment regarding these developments. The situation underscores the difficulty companies face in preventing knowledge leakage without stifling legitimate research.
* DeepSeek and Moonshot AI used Astra for their latest model training
* Anthropic claims its new methods cut data extraction by 90 to 99 percent
* OpenAI reported a coordinated campaign involving Moonshot-linked individuals
