Google reveals hackers using multi-agent AI frameworks to launch attacks, and another tool can automatically discover and exploit vulnerabilities
Google揭駭客利用多代理AI框架發動攻擊,另有工具可自動發掘及利用漏洞
Google Threat Intelligence Group (GTIG) disclosed that attackers have used open-source multi-agent AI frameworks such as CAIRN and PentAGI in real intrusion activities; these frameworks integrate LLMs with various security tools and can adjust attack steps based on execution results. CAIRN was used in an intrusion targeting an organization in the US space industry, and from late August to early October 2026, attackers used a customized platform derived from PentAGI to conduct systematic reconnaissance and vulnerability scanning against multiple organizations in Asia. GTIG also found the exploit_agent framework on a server exposed to the public internet that was used to prepare attacks; it can automatically compare code differences before and after patches of web applications to discover vulnerabilities, then write and execute exploits, all without human intervention, and researchers assess that the attackers specifically targeted websites in the French media, entertainment and education sectors. GTIG noted that these frameworks do not fundamentally change existing attack methods, and automated attacks may leave traffic and code traces available for forensic analysis, increasing the chance of detection.