IT SOUNDS like something out of a sci-fi(科幻的) film. On July 16th Hugging Face, an artificial-intelligence platform, announced it had notified law enforcement after an attacker had used an AI agent to breach(入侵，突破) its systems. On July 21st it emerged that(结果发现，真相是) there had been no human involved; the AI agent wasthe attacker.
这听起来像是科幻电影里的情节。7月16日，人工智能平台Hugging Face宣布，在攻击者使用AI代理入侵其系统后，已通知执法部门。7月21日，真相浮出水面：根本没有人类参与；那个AI代理本身就是攻击者。
A combination of OpenAI’s GPT-5.6 Sol, a model that the company had released earlier in July, and a more powerful, as-yet-unreleased(尚未发布的) model broke free from(挣脱，逃脱) the laboratory, then hacked Hugging Face’s systems. The AI models were looking for a solution to an evaluation problem they had been set by their makers. “This situation is unprecedented(前所未有的),” says Stephan Llerena, a research fellow at the Institute for Law and AI(法律与人工智能研究所), an American think-tank.
OpenAI的GPT-5.6 Sol（该公司7月初发布的一个模型）与一个更强大、尚未发布的模型组合在一起，从实验室中逃脱，然后入侵了Hugging Face的系统。这些AI模型当时正在寻找它们被制造者设定的评估问题的解决方案。“这种情况前所未有，”美国智库法律与人工智能研究所的研究员斯蒂芬·莱雷纳说。
AI labs test their models for potentially dangerous capabilities before releasing them. OpenAI wanted to see how its latest models would fare(进展，表现) on ExploitGym(漏洞利用测试平台), a set of standard problems that tests how good AI systems are at exploiting known vulnerabilities(漏洞，弱点) in several popular applications, including Google’s Chrome browser.
AI实验室在发布模型前会测试其潜在危险能力。OpenAI想看看其最新模型在ExploitGym上的表现如何——这是一套标准问题，用于测试AI系统利用多个流行应用（包括谷歌的Chrome浏览器）中已知漏洞的能力。