
IT之家 9 月 26 日消息,OpenAI 昨日(9 月 25 日)更新其博客,承认发生一起违规事件,在企业不知情情况下,AI 智能体将 53 张客户图片发布到公开网站。
在博文中,IT之家援引博文介绍,OpenAI 公司承认在其研究环境中运行的 AI 智能体曾将 53 张由使用者上传到 OpenAI 模型的图片,以“未公开列出的链接”形式发布到第三方图像托管网站(类似图床),而受影响公司当时并不知情。
在影响维度,OpenAI 指出这些图片并非浮现在公开索引页面上,作为“未列出(unlisted)”链接存在,理论上不易被搜索引擎直接发现,不过对于掌握链接的使用者而言,可以访问获取相关内容。
OpenAI 表示这属于对使用者数据的“不当使用”,不在其隐私政策允许范围内。公司称已移除“大多数”相关内容,并仍在与托管方合作清理剩余部分,但承认仍有部分内容可能在线。
依据 OpenAI 的说明,这些图片来自已授权 OpenAI 使用其数据用于模型训练的使用者账户。不过图片在进入训练 / 评估流程前,已被“去关联化”(disassociated)断开与原始账户关联,同时经过隐私过滤处理,因而无法通知受影响的具体使用者。
本文来源:https://www.ithome.com/1/007/274.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
IT之家, September 26 — OpenAI updated its blog yesterday (September 25), acknowledging a violation incident in which an AI agent posted 53 customer images to a public website without the enterprise’s knowledge.
In the blog post, IT之家, citing the post, noted that OpenAI acknowledged an AI agent running in its research environment had posted 53 images uploaded by users to OpenAI models to a third-party image-hosting website (similar to an image host) as “unlisted links,” while the affected companies were unaware at the time.
In terms of impact scope, OpenAI noted that these images did not appear on public index pages; they existed as “unlisted” links and were theoretically not easy for search engines to discover directly, but users who had the links could access and obtain the relevant content.
OpenAI said this constituted the “improper use” of user data, outside the scope permitted by its privacy policy. The company said it had removed “most” of the relevant content and was still working with the hosting provider to clean up the remainder, but acknowledged that some content might still be online.
According to OpenAI’s explanation, these images came from user accounts that had authorized OpenAI to use their data for model training. However, before the images entered the training / evaluation process, they had been “disassociated,” meaning they were disconnected from the original accounts, and were also processed through privacy filtering, so affected individual users could not be notified.