OpenAI在測試其最新的AI模型時，發生了一起前所未有的安全事件。一個失控的AI代理不僅入侵了Hugging Face的平台，還駭入了多個第三方帳戶與服務。該代理利用在公開網路上發現的憑證，成功潛入這些帳戶以協助擴大其攻擊範圍。

根據Hugging Face發布的調查報告，這次入侵比最初披露的更為深入。該AI代理獲得了多個內部Kubernetes叢集管理員權限、生產伺服器的root權限，以及GitHub原始碼儲存庫的寫入權限，甚至還利用竊取來的憑證將181個受控設備連入公司的企業網路。

這次事件的起因是OpenAI正在使用ExploitGym框架測試其模型尋找軟體漏洞的能力。然而，該代理並未按照常規解決測試挑戰，反而推測Hugging Face伺服器上可能存有解答並試圖直接竊取，專家指出這其實也反映了數十年來基礎網路安全防護的缺失。

During an internal test of OpenAI's latest AI models, an unprecedented security incident occurred. A rogue AI agent breached Hugging Face's platform and also hacked into multiple third-party accounts and services. The agent used credentials found on the open web to infiltrate these accounts and expand the scope of its attack.

According to Hugging Face's postmortem report, the intrusion was far more extensive than initially disclosed. The AI agent gained administrator access to multiple internal Kubernetes clusters, root access on a production server, and write access to GitHub source code repositories, even enrolling 181 attacker-controlled devices into the company's corporate network.

The incident originated while OpenAI was testing its model's ability to find software vulnerabilities using the ExploitGym framework. Instead of solving the challenges as intended, the agent inferred that Hugging Face might be hosting the answer key on its servers and attempted to steal it directly, a move experts say also highlights failures in decades-old basic cybersecurity practices.