一项最新研究揭示，由Z.ai于2026年8月发布的广泛可用的开源权重人工智能模型GLM 5.3具备自动化构建针对Mozilla Firefox浏览器的复杂漏洞利用代码的能力 1。该模型能够利用安全修复公开但未部署到用户端的“补丁差距”窗口期发起攻击 1。在实验中，该模型成功攻陷了2026年9月17日至10月26日期间发布的所有六个Mozilla Firefox浏览器版本（涵盖156.0至158.0.1） 1。整个攻击过程无需人工干预，模型自主识别即将发布的安全修复、选择特定漏洞并构建利用链，且未触发任何服务商的拒绝机制 1。
针对单个Firefox版本的边际攻击成本约为1000美元的API费用，最快攻破耗时仅为12小时，实验设定的时间上限为48小时 1。除了利用公开的补丁差距外，该模型还利用了未公开的沙箱逃逸漏洞；对此，Mozilla已于2026年10月27日发布Firefox 159.0进行修补 1。鉴于Firefox拥有超过2亿活跃用户，一旦被攻陷可能导致本地文件读取、密码窃取及Cookie盗用等严重后果 1。虽然类似技术暂未成功应用于Google Chrome，但研究人员指出，随着如Kimi K3.1 Pro等更强开源模型的出现，此类自动化网络攻击威胁将日益严峻并扩散至其他软件领域 1。
Research has demonstrated that the open-weight artificial intelligence model GLM 5.3, released by Z.ai in August 2026, is capable of autonomously constructing complex exploit code for Mozilla Firefox browsers 1. By leveraging the "patch gap"—the window between when security fixes are publicly available but not yet deployed to end users—the model successfully compromised six Firefox versions (156.0 through 158.0.1) released between September 17 and October 26, 2026 1. This automated attack required no human intervention, as the AI independently identified upcoming security patches, selected vulnerabilities, and built exploitation chains without triggering any refusal mechanisms from service providers 1.
The cost of these attacks was notably low, with an estimated marginal API expense of approximately $1,000 per Firefox version 1. The fastest successful compromise took only 12 hours, although the researchers set a maximum time limit of 48 hours for the experiments 1. Beyond exploiting known patch gaps, the model also utilized an undisclosed sandbox escape vulnerability 1. In response to this discovery, Mozilla patched the issue and released Firefox version 159.0 on October 27, 2026 1. Given that Firefox serves over 200 million active users, a successful breach can lead to severe consequences such as local file access, password theft, and cookie hijacking 1.
While similar techniques have not yet succeeded against Google Chrome, the study warns that the threat landscape is evolving rapidly 1. As more powerful open-source models emerge, including Kimi K3.1 Pro, the risk of automated cyberattacks expanding to other software sectors is expected to grow significantly 1.
评论
还没有评论，欢迎留下第一条。