Anthropic 收紧防御后，OpenAI GPT-6 Astra 成为中国 AI 实验室蒸馏的主要目标
OpenAI's Astra becomes the main target for Chinese AI labs after Anthropic tightens its defenses
- 1据 The Information 报道，DeepSeek 和 Moonshot AI 转而大量利用 OpenAI 最强模型 GPT-6 Astra 的输出训练最新模型
- 2知情人士称涉事模型包括 Kimi K3.1 Pro 和 V4.1-Flash 1017，Moonshot 拒绝置评，DeepSeek 未回应
- 3Anthropic 称其新防御方法将攻击者可提取的数据量削减 90% 至 99%，此前 Claude 是蒸馏的首选目标
- 4OpenAI 称 9 月底发现一起与 Moonshot 关联人员相关的协同蒸馏行动，并表示攻击者会随每次新防御调整手法
蒸馏与反蒸馏是持续博弈，Anthropic 收紧防御并未终结蒸馏，只是把目标从 Claude 推向了 OpenAI 的 Astra。
- 影响：模型能力宣传与实际训练来源之间可能存在差距，选型时值得把能力来源可信度纳入评估；厂商禁止蒸馏的条款并未完全拦住行为，头部模型间的能力差距和更新节奏会持续受这场攻防影响。
- 参考：可考虑在模型选型评估中增加训练来源与合规问询，要求供应商说明能力数据的口径；值得把各厂商蒸馏防御公告的更新频率，作为评估其安全能力的参考之一。
- 依据：核心事实援引 The Information 及知情人士，DeepSeek 与 Moonshot 均未回应置评；90% 至 99% 的削减幅度为 Anthropic 官方口径，暂无独立验证。
Chinese AI labs are now mainly tapping OpenAI's most powerful model, GPT-6 Astra, instead of Anthropic's Claude. According to The Information, DeepSeek and Moonshot AI leaned heavily on Astra's outputs to train their latest models. People familiar with the matter say this includes Kimi K3.1 Pro and V4.1-Flash 1017.
This kind of distillation only works if the labs get the models to reveal their hidden reasoning steps. OpenAI and Anthropic both ban that. Moonshot declined to comment, and DeepSeek didn't respond to a request for comment. Claude used to be the go-to target, but Anthropic has tightened its defenses. The company says its new methods cut the amount of data attackers can pull by 90 to 99 percent.
In late September, OpenAI reported a coordinated distillation campaign that it says involved people linked to Moonshot. An OpenAI spokesperson said the company takes the problem very seriously, but until now attackers adjusted their tactics every time it rolled out a new defense.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
来源：The Decoder：AI News · the-decoder.com