AI Security AI安全 6h ago Updated 2h ago 更新于 2小时前 49

Nine-Year-Old RefluXFS Linux Flaw Gives Local Users Root on Default RHEL Installs 九年历史的RefluXFS Linux漏洞使默认RHEL安装中的本地用户获得Root权限

RefluXFS (CVE-2026-64600) is a critical local privilege escalation flaw in the Linux kernel's XFS filesystem that allows unprivileged users to overwrite root-owned files and gain persistent root access. The vulnerability stems from a race condition in copy-on-write reflink handling, where stale block mappings are used after a lock cycle, enabling direct I/O writes to bypass inode checks. Default installations of RHEL 8/9/10, Fedora Server, Amazon Linux, and various derivatives are exposed if XFS RefluXFS (CVE-2026-64600) 是 Linux XFS 文件系统中存在九年的严重漏洞,允许未授权本地用户通过竞态条件覆盖 root 拥有的文件并获取持久性 root 权限。 该漏洞利用 XFS 的 reflink(写时复制)机制和内核锁循环中的陈旧映射错误,在 RHEL、Fedora Server 和 Amazon Linux 等默认配置下即可被利用。 漏洞由 Qualys 使用 Anthropic 的 Claude Mythos Preview AI 模型发现,AI 成功定位了竞态条件、编写了利用代码并起草了安全公告。 补丁已于 7 月 16 日合并入上游内核,主要发行版

75
Hot 热度
65
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • RefluXFS (CVE-2026-64600) is a critical local privilege escalation flaw in the Linux kernel's XFS filesystem that allows unprivileged users to overwrite root-owned files and gain persistent root access.
  • The vulnerability stems from a race condition in copy-on-write reflink handling, where stale block mappings are used after a lock cycle, enabling direct I/O writes to bypass inode checks.
  • Default installations of RHEL 8/9/10, Fedora Server, Amazon Linux, and various derivatives are exposed if XFS is mounted with reflink=1, a common default configuration.
  • The bug was discovered by an AI model (Claude Mythos Preview) tasked with finding vulnerabilities similar to Dirty COW, highlighting the emerging role of AI in security research.
  • Patches have been merged upstream and distributed by vendors; systems must update their kernels and reboot to mitigate the risk, particularly on multi-tenant or shared-hosting environments.

Why It Matters

This vulnerability is significant because it affects widely deployed enterprise Linux distributions by default, turning a standard filesystem feature into a critical security hole for any user with local shell access. It demonstrates a new paradigm in vulnerability discovery where frontier AI models can identify complex kernel race conditions that human researchers might overlook, potentially accelerating the pace of security auditing. For practitioners, it underscores the immediate need to verify patch status on XFS volumes, especially in cloud and containerized environments where local user isolation is paramount.

Technical Details

  • Root Cause: A check-then-use error in xfs_reflink_fill_cow_hole() and xfs_reflink_fill_delalloc(). The kernel captures a data-fork mapping under the inode lock, releases the lock to reserve transaction space, and then reacquires it. During this gap, another writer can remap the cloned file to new blocks. When the first writer resumes, it uses the stale mapping, pointing to blocks owned only by the original protected file.
  • Exploitation Mechanism: An attacker clones a root-owned file using FICLONE (requiring only read access), then races concurrent O_DIRECT writes against the clone. Because Direct I/O bypasses the page cache and lacks revalidation hooks, the write lands directly on disk at the stale address, overwriting the target file without changing its metadata (ownership, permissions, timestamps).
  • Affected Configurations: Linux kernels 4.11 and later. Specifically affected are systems running Red Hat Enterprise Linux (8, 9, 10), CentOS Stream, Oracle Linux, Rocky Linux, AlmaLinux, CloudLinux, Fedora Server (31+), and Amazon Linux (2023 and specific Amazon Linux 2 images). Debian/Ubuntu are generally safe unless XFS with reflink is manually configured.
  • AI Discovery: Qualys used Anthropic's Claude Mythos Preview to find the flaw. The model identified the race condition, wrote a working exploit, and drafted the advisory, which researchers then reproduced and coordinated for disclosure.
  • Mitigation: Vendors have released backported kernels. Users must apply the relevant errata (e.g., RHSA-2026:39179 for RHEL 8) and reboot. Verification can be done via xfs_info / | grep reflink=.

Industry Insight

  • AI-Augmented Security Research: The successful use of AI to discover a nine-year-old kernel bug suggests that AI tools will become standard components in vulnerability hunting pipelines. Organizations should consider integrating AI-assisted code analysis and fuzzing into their internal security teams' workflows.
  • Default Configuration Risks: The fact that reflink=1 is enabled by default on many major distros highlights the danger of assuming secure defaults. Security audits must explicitly verify filesystem mount options and kernel versions, not just application-level configurations.
  • Patch Velocity Expectations: With AI potentially shortening the time between bug discovery and exploit availability, the window for remediation may shrink. Automated patch management and continuous compliance scanning for kernel updates should be prioritized, especially for multi-tenant infrastructure where a single compromised user account could lead to full system takeover.

TL;DR

  • RefluXFS (CVE-2026-64600) 是 Linux XFS 文件系统中存在九年的严重漏洞,允许未授权本地用户通过竞态条件覆盖 root 拥有的文件并获取持久性 root 权限。
  • 该漏洞利用 XFS 的 reflink(写时复制)机制和内核锁循环中的陈旧映射错误,在 RHEL、Fedora Server 和 Amazon Linux 等默认配置下即可被利用。
  • 漏洞由 Qualys 使用 Anthropic 的 Claude Mythos Preview AI 模型发现,AI 成功定位了竞态条件、编写了利用代码并起草了安全公告。
  • 补丁已于 7 月 16 日合并入上游内核,主要发行版如 Red Hat 和 Debian 已开始分发修复后的内核版本,建议受影响用户立即升级。

为什么值得看

这篇文章展示了 AI 在底层系统安全研究中的突破性应用,证明了前沿大模型能够独立发现复杂的内核竞态条件漏洞。同时,它揭示了长期存在的 XFS 文件系统设计缺陷对主流企业级 Linux 发行版的广泛影响,为运维团队提供了紧急的安全加固指引。

技术解析

  • 漏洞原理:利用 FICLONE 创建 root 文件的克隆,并通过并发 O_DIRECT 写入触发竞态。在内核释放 inode 锁以预留事务空间期间,第二个写入者完成写时复制操作并重新映射块;当第一个写入者重新获取锁时,其使用的数据分叉映射已过期(stale),导致写入操作意外覆盖原始受保护文件。
  • 利用条件:系统运行 Linux 4.11 及以上版本且未打补丁;XFS 文件系统启用 reflink=1;攻击者可写的目录与目标受保护文件位于同一 XFS 文件系统上。
  • 隐蔽性与持久性:由于 Direct I/O 绕过页面缓存且无验证钩子,写入直接落盘且不修改目标 inode 元数据,因此不会留下内核警告或日志记录。即使重启,被篡改的二进制文件仍保留 setuid 位,确保持久 root 访问。
  • 修复方案:补丁通过在内核释放锁前快照 ip->i_df.if_seq 计数器,并在重新获取锁后使用 xfs_bmapi_read() 重新读取数据分叉来修复陈旧映射问题,涉及 xfs_reflink_fill_cow_hole()xfs_reflink_fill_delalloc() 两个辅助函数。

行业启示

  • AI 驱动的安全研究范式转变:Qualys 成功利用 AI 发现长达九年的内核漏洞,标志着 AI 在自动化漏洞挖掘和复杂逻辑分析方面已从概念验证走向实战能力,安全团队应重视并整合 AI 工具以提升代码审计效率。
  • 基础设施供应链风险:XFS 作为许多主流 Linux 发行版的默认文件系统,其底层缺陷影响了 RHEL、Fedora 和 Amazon Linux 等广泛生态,强调了定期审查内核更新和文件系统配置的重要性,特别是针对多租户环境。
  • 应急响应优先级:鉴于漏洞利用窗口短且无需特殊权限,企业和云服务商应立即检查是否启用了 reflink=1,并优先修补暴露于不可信代码(如 CI/CD 作业、共享主机)的系统,以消除本地提权风险。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Open Source 开源 Security 安全