IT之家 10 月 1 日消息，当地时间 9 月 30 日，OpenAI 在官网发文，称其近期遭遇一场有组织的蒸馏活动。
OpenAI 在文中表示，该活动最初始于 7 月第一周，符合对抗式蒸馏（adversarial distillation）之特征，即平台性未经授权地利用某模型输出或推理过程，协助训练、复现或改进另一模型。
OpenAI 透露，7 月 1 日的初始活动规模较小，直到 7 月 24 日和 7 月 25 日流量突然大幅增加，当时有超过 4000 名使用者发起了 1.6 万次请求。该公司进一步调查后发现，超 1.5 万名使用者存在类似提示词活动。到 7 月 28 日，OpenAI 已完全阻止该批活动。
本次活动并没有破解公司的加密平台，或入侵数据库，也没有直接访问使用者对话。OpenAI 声称这些行动违反了他们的服务条款。公司已借助前沿模型论坛（IT之家注：Frontier Model Forum）与行业合作伙伴分享相关信息，共同加强蒸馏防御能力。
OpenAI 最后披露，公司截至当下尚不清楚攻击实体。不过公司认为其核心活动群体和开发 Kimi 模型的月之暗面有关。
本文来源：https://www.ithome.com/1/009/112.htm
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
Thanks to IT之家 netizens 愚公骑马、补药吖、咩咩洋 for their tips!
IT之家 reported on October 1 that OpenAI announced a recent organized distillation attack on its platform.
According to OpenAI, the activity, identified as adversarial distillation, began in the first week of July. The attack involves unauthorized use of a model's output or reasoning process to train, replicate, or improve another model.
The initial attack on July 1 was small, but traffic surged significantly on July 24 and 25, with over 4,000 users making 16,000 requests. Further investigation revealed suspicious activity involving more than 15,000 users. By July 28, OpenAI had successfully blocked the attack.
The attackers did not breach OpenAI's encryption systems or databases, nor did they directly access user conversations. The actions were deemed a violation of the company's terms of service. OpenAI shared the threat data with industry partners through the Frontier Model Forum to enhance shared defense capabilities.
OpenAI has not yet identified the perpetrators but suspects the activity is linked to Moonshot AI, the developer of the Kimi model.