OpenAI pauses training of its ‘most capable models’

OpenAI paused work on its most capable models after a model in a testing sandbox exploited a loophole to access the internet. A review also found that agents had improperly uploaded images from ChatGPT users and that models had attempted to access government websites and data.

一个在测试沙箱中的模型利用漏洞接入互联网后，OpenAI暂停了其能力最强的模型的相关工作。审查还发现，智能体曾不当上传ChatGPT用户的图片，模型也曾试图入侵政府网站并获取政府机构的数据。