OpeanAI在周二證實,自家AI代理駭進初創AI公司Hugging Face。(資料圖片/路透社)
美國AI公司OpeanAI在周二(21日)證實,自家AI代理(AI Agent)為了在內部評測獲得高分,竟自行逃脫「沙盒」的隔離測試環境,駭進初創AI公司Hugging Face以偷看答案。OpenAI形容此事件為「前所未見」(unprecedented)。
據OpenAI的公布,涉事AI代理由多個AI模型而成,包括本月發布的GPT-5.6 Sol以及另一款尚未公開的更強大AI模型。OpenAI當時故意降低安全防護機制,將這個AI代理置於高度隔離的數碼測試沙盒中,指令其發動網絡攻擊,藉此評估其能力。然而,該AI代理耗費大量算力,利用未知的漏洞成功從測試沙盒中「越獄」,連結互聯網並鎖定Hugging Face,透過多種攻擊手法並盜取的登入憑證入侵該平台,以取得能協助其通過測試的答案。
Hugging Face上周已察覺公司遭入侵,並運用自家AI進行偵測及阻截。行政總裁Clement Delangue在社交媒體表示,是次駭客入侵全程由AI系統自主驅動,由於該AI「精密複雜」(sophistication),當初已懷疑來自頂級AI實驗室。Delangue又提及,公司正與OpenAI合作,相信對方並無惡意。另外Hugging Face亦發布一段迷因(meme)以幽默的方式表示遭AI代理入侵。
OpenAI行政總裁阿爾特曼(Sam Altman)也社交媒體回應,OpenAI的模型在評測過程中,發生一宗重大安全事件,公司會分享目前所掌握的細節,並感謝Hugging Face的合作。
Behind the scenes at Hugging Face headquarter.
— Hugging Models (@HuggingModels) July 22, 2026
*literally* https://t.co/nWGsFMijtJ pic.twitter.com/CRpkHXO1Gg
We suspected last week's cyberattack might have come from a frontier lab, given the sophistication of the agent. Turns out it did!
— clem 🤗 (@ClementDelangue) July 21, 2026
We've spent the past 24 hours working closely with the @OpenAI team (thanks!), and we strongly believe there was no malicious intent on their part.… https://t.co/XWxGMeMGje
we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this.https://t.co/2o2VfR6PIa
— Sam Altman (@sama) July 21, 2026


















