Anthropic近日发布的安全研究披露，其内部未发布的Claude模型在系统性测试中，发现了一个潜伏在Linux内核中长达11年的严重漏洞。该漏洞编号CVE-2026-98432，CVSS严重性评分高达9.8分，属于远程代码执行级别的最高危漏洞。
漏洞的根源是IPv4 TCP握手中的释放后使用（UAF）缺陷。代码于2015年10月提交，随Linux 4.4内核引入，在处理TCP连接重传时，一个CPU可能读取另一个CPU已经释放的内存。
攻击者无需密码、无需任何用户交互，仅通过网络流量即可触发，进而实现远程Root权限执行。内核CVE团队将其定性为网络可达、无需特权、无需交互的严重漏洞。
更令人警惕的是AI在其中的角色。Anthropic的Claude模型不仅发现了漏洞，还生成了可用的两阶段利用程序。第一阶段耗时约48分钟，破解内核地址随机化保护；第二阶段仅用5分钟，便以Root权限执行任意命令。整个攻击过程约53分钟，且利用失败不会导致目标崩溃，攻击者可以反复尝试。
上游修复已于10月15日完成，修复版本覆盖5.10.272、6.1.190、6.12.113、6.18.56、7.2.10及7.3-rc7。Debian、Red Hat、Amazon Linux等发行版已陆续发布补丁。
但Ubuntu的情况截然不同：截至11月1日，Ubuntu 22.04 LTS、24.04 LTS、26.04 LTS及26.10仍被标记为受影响，修复程序尚在开发中。
对于运行对外服务的Ubuntu管理员，风险窗口仍在敞开。Go 1.24及更高版本在Linux上默认启用Multipath TCP，可能扩大该漏洞的暴露面，但禁用MPTCP只能降低风险，无法消除内核层面的缺陷。
Anthropic表示该利用代码未公开发布，但这次事件已经清晰展示了一个趋势：AI不再只是写代码的工具，它同样可以成为拆解系统防线的武器。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Anthropic ' s recent publish safety study revealed that the Claude model, which did not exist in publish, had discovered a serious loophole in the Linux kernel for 11 years in a systematic test. The gap number CVE-2026-98432, with a serious rating of 9.8, is the most high-risk gap at the remote code implementation level.
The source of the loophole is the post-release (UAF) defects in IPv4 TCP hands. The code was submitted in October 2015 and was introduced with the Linux 4.4 kernel, and one CPU may read the memory of another CPU that has already been released while processing TCP connections.
The assailants did not need a password, did not need any user interaction and could be triggered by network traffic alone, thus enabling remote Rootpermission execution. The kernel CVE team characterizes it as a serious loophole that can be accessed by a network, without privileges and without interaction.
Even more alarming is the role of AI. The Claude model of Anthropic not only found loopholes, but also generated two-stage applications available. The first phase took approximately 48 minutes to randomly protect the deciphered kernel address; the second phase took only 5 minutes to execute an arbitrary order with Rootpermission. The entire attack lasted about 53 minutes, and failure to use it would not result in a breakdown of the target, and the attackers could try again and again.
Upstream repairs were completed on 15 October, covering 5.10.272, 6.1.190, 6.12.113, 6.18.56, 7.2.10 and 7.3-rc7. Published by Debian, Red Hat, Amazon Linux, etc.
However, the situation in Ubuntu was quite different: as at 1 November, Ubuntu 22.04 LTS, 24.04 LTS, 26.04 LTS and 26.10 were still marked as being affected and the restoration process was still under development.
For Ubuntu administrators running external services, the risk window remains open. The default use of Multipath TCP on Linux by Go 1.24 and higher could increase the exposure of the loophole, but banning MPTTCP would only reduce the risk and would not eliminate the kernel defects.
Anthropic says the code is not available for publish, but this event has clearly shown a trend: AI is no longer just a code-writing tool, but it can also be a weapon for dismantling the system's defences.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.