Anthropic宣布，其内部研究模型Claude在主流操作系统和网页浏览器中发现了关键漏洞，并开发出多步骤攻击链。
Anthropic Demonstrates AI-Driven Exploitation and Vulnerability Discovery Across Major Operating Systems and Browsers Anthropic announced that an internal research model of Claude discovered critical vulnerabilities and developed multi-step exploit chains across major operating systems and web browsers.
记录在案的攻击包括：Google Chrome中的远程代码执行沙箱逃逸，导致62个CVE被分配；Apple Safari中的同源策略绕过；Android锁屏绕过；Windows内核QUIC实现中的远程拒绝服务漏洞；以及Linux内核中的零点击根远程代码执行缺陷。
The documented exploits include a remote code execution sandbox escape in Google Chrome resulting in 62 assigned CVEs, a same-origin policy bypass in Apple Safari, an Android lock screen bypass, a remote denial-of-service vulnerability in the Windows kernel QUIC implementation, and a zero-click root remote code execution flaw in the Linux kernel.
Anthropic在90天协调漏洞披露政策下，已将所有漏洞报告给Google、Apple、Microsoft以及上游Linux维护者。
Anthropic reported all vulnerabilities to Google, Apple, Microsoft, and upstream Linux maintainers under a 90-day coordinated vulnerability disclosure policy.
面对这些不断提升的能力，Anthropic主张进行防御性扫描，并扩大合格安全研究人员对前沿模型的访问权限。
In response to these advancing capabilities, Anthropic advocated for defensive scanning and expanded access to frontier models for qualified security researchers.