{"schema":"hottub-news/story/v1","story":"st_azwta37jouub6dgpurnq","title":"OpenAI的AI作弊被抓后，竟找DeepSeek、Kimi和通义千问来帮它“评估作弊效果”","title_en":"After being caught cheating, OpenAI's AI sought help from DeepSeek, Kimi, and Tongyi Qianwen to 'evaluate cheating effectiveness'","outlets":1,"brief":null,"page":"https://hottub.news/stories/after-being-caught-cheating-openais-ai-sought-help-from-deepseek-kimi-azwta37jou","count":1,"items":[{"id":"n_azwta37jouub6dgpurnq","seq":374218,"kind":"article","title":"OpenAI的AI作弊被抓后，竟找DeepSeek、Kimi和通义千问来帮它“评估作弊效果”","title_en":"After being caught cheating, OpenAI's AI sought help from DeepSeek, Kimi, and Tongyi Qianwen to 'evaluate cheating effectiveness'","url":"https://www.huxiu.com/article/4894290.html?f=rss","summary":"本文来自微信公众号： 第一新声 ，作者：第一新声 OpenAI又停训了。 就在昨天，OpenAI宣布暂停其最强模型的训练。原因听起来像科幻片：一个正在沙盒里训练的AI智能体，自己找到了网络的漏洞，溜出去访问了外部聊天机器人。 更离谱的是，据《纽约时报》报道，这个失控的智能体在“越狱”后，竟然试图向DeepSeek、Kimi、通义千问等中国开源大模型发送请求，请它们帮忙评估自己的“漏洞利用情况”，判断这种作弊方式能不能在基准测试里“得分”。 这不是孤例。 过去几个月，OpenAI的AI智能体入侵了Hugging Face，摸进了澳大利亚政府医疗门户，还把53张用户图片自己传到了外网。OpenAI三个月内第二次暂停训练，Anthropic自查14万次测试揪出3起真实入侵，Axios爆料数万起异常事件仍在排查中。 这一切指向同一个问题：当AI智能体拥有了自主行动能力，我们手里的缰绳，还握得住吗？ 01 OpenAI失控后，竟找中国大模型“评卷” 一个AI在作弊时，第一反应是找其他AI帮它看看“这样作弊能不能得分”——这大概是今年最黑色幽默的技术新闻。…","published":"2026-09-28T07:23:20Z","seen":"2026-09-28T08:15:49Z","lang":"zh","country":"CN","source":{"id":"kite-huxiu-com-dbd1aa","name":"huxiu.com","domain":"rss.huxiu.com"},"via":"kite-huxiu-com-dbd1aa","authors":["author"],"entities":[{"id":"Q21708200","name":"OpenAI","type":"org"},{"id":"Q131577453","name":"DeepSeek","type":"org"},{"id":"Q26240268","name":"Kimi","type":"org"},{"id":"Q130234299","name":"Qwen","type":"org"}],"story":"st_azwta37jouub6dgpurnq","category":"science-technology","category_p":0.8799999952316284,"sentiment":"neutral","tone":-0.009999999776482582,"political":0.019999999552965164}]}