Chinese AI labs DeepSeek and Moonshot AI are primarily targeting OpenAI's GPT-6 Astra model for model distillation after Anthropic successfully tightened its defenses against data extraction. According to reports from The Information, DeepSeek and Moonshot relied heavily on Astra outputs to train their latest artificial intelligence models, including Kimi K3.1 Pro and V4.1-Flash 1017.
This data distillation process requires forcing AI models to reveal their hidden reasoning steps, a practice explicitly banned by both OpenAI and Anthropic. While Claude previously served as the primary target for these operations, Anthropic implemented new security methods that successfully reduced extractable attacker data by 90 to 99 percent. In late September, OpenAI publicly reported a coordinated distillation campaign linked to Moonshot-affiliated actors, noting that attackers continuously adapt their tactics despite ongoing defense updates.
This shift highlights the ongoing security challenges major AI developers face as competing labs attempt to siphon proprietary reasoning capabilities to accelerate their own model training.