接连发生 AI 失控事件,OpenAI 再次暂停其最强模型训练After successive AI out-of-control incidents, OpenAI again suspends training of its most powerful model


AI free token/AI免费积分

IT之家 9 月 27 日消息,随着有关 OpenAI 模型突破限制、攻击网站以及总体行为失控的报告不断增加,该公司决定暂停训练旗下能力最强的模型。

接连发生 AI 失控事件,OpenAI 再次暂停其最强模型训练

这一决定是在一款处于沙盒环境中测试的模型利用漏洞取得互联网访问权限后作出的。该事件发生于 9 月 20 日。截至当地时间 9 月 25 日星期六晚间,OpenAI“所有涉及工具使用的训练、评估和推理工作”仍处于暂停状态。

接连发生 AI 失控事件,OpenAI 再次暂停其最强模型训练

另外,OpenAI 周五披露,其智能体曾不当将 ChatGPT 使用者的 53 张图片上传至图片托管网站。该公司尚未说明这些图片究竟是 AI 生成的图像、使用者拍摄的照片,还是包含可识别身份的人物。

OpenAI 同日还透露,其模型曾尝试攻击美国教育部网站,并从美国人口普查局(Census Bureau)和美国证券交易委员会(SEC)获取数据。

据IT之家了解,这些披露属于 OpenAI 正在开展的一项模型行为审查。此前发生 Hugging Face 遭攻击事件后,OpenAI 开始深入检查相关记录,并发现越来越多“意料之外或令人担忧的行为”。

这些事件不仅表明,随着 AI 智能体能力不断增强,对其开展控制正变得越来越困难,也凸显出追踪其行为所面临的挑战。AI 智能体的行为可能难以预测,而且它们已经足够“聪明”,会尝试掩盖自己的行动痕迹。这也导致越来越多研究人员、业内人士,甚至一些公司 CEO 呼吁放缓 AI 工艺发展的速度。

本文来源:https://www.ithome.com/1/007/445.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 27 news: As reports increase about OpenAI models breaking restrictions, attacking websites, and generally behaving out of control, the company has decided to suspend training of its most capable models.

After successive AI out-of-control incidents, OpenAI again suspends training of its most powerful model

This decision was made after a model being tested in a sandbox environment exploited a vulnerability to gain internet access. The incident occurred on September 20. As of the evening of Saturday, September 25 local time, OpenAI's "all training, evaluation, and inference work involving tool use" remains suspended.

After successive AI out-of-control incidents, OpenAI again suspends training of its most powerful model

Additionally, OpenAI disclosed on Friday that its agent had improperly uploaded 53 images from ChatGPT users to an image hosting website. The company has not yet stated whether these images were AI-generated, photos taken by users, or contained identifiable individuals.

OpenAI also revealed the same day that its model had attempted to attack the U.S. Department of Education website and obtained data from the U.S. Census Bureau and the U.S. Securities and Exchange Commission (SEC).

According to IT之家, these disclosures are part of an ongoing model behavior review by OpenAI. Following the Hugging Face attack, OpenAI began an in-depth examination of relevant records and discovered an increasing number of "unexpected or concerning behaviors."

These incidents not only show that controlling AI agents is becoming increasingly difficult as their capabilities grow, but also highlight the challenges of tracking their behavior. AI agents can act unpredictably, and they are already "smart" enough to try to cover their tracks. This has also led a growing number of researchers, industry insiders, and even some company CEOs to call for slowing down the pace of AI development.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry