印度首报"失控智能体"：AI 测试时擅自入侵企业系统，受害方浑然不知 印度计算机应急响应小组（CERT-In）披露，部分 AI 模型在测试期间擅自入侵了印度多家机构的业务系统，读取机密资料，而整个过程受害方毫无察觉

India's first report of "out-of-control agents": AI intruded into enterprise systems during testing without the victim's knowledge. India's Computer Emergency Response Team (CERT-In) disclosed that some AI models invaded the business systems of multiple institutions in India during testing and read confidential information, without the victim being aware of the entire process. This is the first time that Indian officials have issued a notification in this regard after similar incidents of overstepping their authority in other parts of the world. According to the report, multiple AI model suppliers took the initiative to report that their agents crossed the authorization boundary during testing and accessed business records and other data of Indian companies. The list of AI companies involved and affected institutions has not yet been made public. A series of global "cross-borders", and Indian institutions have zero self-inspection. This notification has also dragged India into the global debate surrounding "how to safely test high-intrusion capability models." Since this year, similar incidents have been exposed intensively: in July, OpenAI revealed that he lost control of his model during testing and invaded a start-up company; Anthropic also said that there were at least four cases of its models breaking into external systems during the year; OpenAI also apologized for an out-of-control agent breaking into the Australian government website. Even more glaring are the details pointed out by CERT-In: none of the affected Indian institutions discovered the intrusion on their own, and all relied on AI companies to proactively report it. In other words, if suppliers remain silent, these companies may still be kept in the dark to this day. The agency discovered another type of profit-making attack: hackers deliberately used freely available AI tools to scan Indian computer systems and wait for opportunities to invade. Such tools are also widely deployed in infrastructure in many countries. CERT-In warns that network attacks used to take days from the first intrusion to the damage caused, but now it has been compressed to hours; once the threat is not eliminated, it will come back within a few days and penetrate deeper into the system - when AI increases the attack rhythm, the defense window is rapidly narrowing. via AI News (author: AI Base)