安全研究员 Mehmet Ince 借助 Claude Opus AI 助手深入分析 PostgreSQL 内部机制,成功发现并披露了一个存在长达22年的类型混淆漏洞 1。该漏洞允许攻击者通过修改复合类型的属性定义触发内存损坏,进而实现远程代码执行 1。此次修复的核心漏洞被命名为 CVE-2026-6460,相关补丁已于 2026年11月9日随新版本发布 1。
Ince 利用名为 'pg_expert' 的 Claude Opus AI Agent 逐步学习数据库内部原理并辅助生成复杂的利用链代码 1。漏洞的基础机制可追溯至 2004 年的 commit 077db40fa1f3,攻击者可在准备查询后使用 ALTER TYPE 更改字段类型(如从 UUID 改为 bytea 或 bigint[]),从而在执行时触发非预期解析导致段错误或任意代码执行 1。具体技术细节包括利用 LANGUAGE internal 功能创建恶意函数,或通过 unnest() 与数组类型混淆泄露内存地址以绕过 ASLR 1。
此次事件影响了所有主流托管 PostgreSQL 服务提供商,包括 AWS RDS/Aurora、Google AlloyDB、Databricks 和 Supabase 等,这些服务商均已收到通报 1。除了主要的 CVE-2026-6460 外,这一发现还关联了52个新 CVE 中的6个直接相关项,其中包括 CVE-2026-120142 1。Ince 于 2026年11月30日在布拉格举行的 PostgreSQL 聚会上发表了演讲,详细记录了从基础数据类型学习到构建复杂利用链的全过程,展示了 AI 在现代网络安全研究中的关键作用 1。
Security researcher Mehmet Ince discovered a critical vulnerability in the PostgreSQL database engine that allows attackers to execute arbitrary code remotely. 1 The flaw is rooted in a type confusion issue involving composite types and Datum structures, having persisted in the codebase for approximately 22 years, with its underlying mechanism traceable to commit 077db40fa1f3 from 2004. 1 By leveraging AI assistance from Claude Opus through an agent named 'pg_expert', Ince analyzed internal mechanisms to construct complex exploitation chains, demonstrating the significant role of artificial intelligence in modern security research. 1
The exploit works by preparing a query and then using the ALTER TYPE command to change field definitions, such as modifying a UUID to bytea or bigint[]. 1 When executed, this triggers unexpected parsing behavior that results in memory corruption, leading to segmentation faults or remote code execution. 1 Ince reported the issue to the PostgreSQL Core Team and multiple cloud service providers, including AWS RDS/Aurora, Google AlloyDB, Databricks, and Supabase. 1 As a result, major patches were released on November 9, 2026, addressing CVE-2026-6460, along with six directly related items among 52 new CVEs identified during the review process. 1 Further technical details revealed that attackers could also use the LANGUAGE internal feature to create malicious functions or employ unnest() with array type confusion to leak memory addresses and bypass ASLR protections. 1 Ince presented his findings at a PostgreSQL gathering in Prague on November 30, 2026. 1
评论
还没有评论,欢迎留下第一条。