Chinese Labs Target GPT-6 Astra After Anthropic's Measures

Chinese laboratories are now favoring GPT-6 Astra for distilling training data, whereas Claude from Anthropic was previously the main target. Anthropic claims to have reduced potential leaks by 90 to 99%, and OpenAI reports a campaign attributed to individuals linked to Moonshot, indicating an ongoing adaptation by attackers.

Strengthened Defenses and Attacker Adaptation According to OpenAI

Anthropic states that it has implemented methods that limit the amount of data attackers can extract from its models by 90 to 99%. At the end of September, OpenAI detected a coordinated distillation campaign attributed to individuals associated with Moonshot. According to OpenAI, attackers modify their tactics each time a new defense is established. An OpenAI spokesperson emphasizes that the company takes this issue seriously. Before this enhancement of protections, Claude was the preferred target of distillation attempts.

DeepSeek and Moonshot Rely on Astra's Outputs

Sources close to the matter report that DeepSeek and Moonshot AI are heavily leveraging the results of GPT-6 Astra to train their recent models, including Kimi K3.1 Pro and V4.1-Flash 1017. This shift in interest makes GPT-6 Astra, touted as OpenAI's most powerful model, the primary target for Chinese laboratories, to the detriment of Claude from Anthropic.

Prohibited Distillation and Scarcity of Public Comments

Distillation is only possible if laboratories manage to obtain models that reveal their hidden reasoning steps, a practice explicitly prohibited by OpenAI and Anthropic. When approached, Moonshot AI declined to comment, while DeepSeek did not respond to requests.