研究团队于2026年10月披露了ChatGPT的一项严重安全缺陷，该漏洞允许攻击者通过发送普通的共享链接，在受害者已登录的会话中执行任意JavaScript代码并劫持用户账户1。由于触发条件仅需受害者打开恶意链接，无需安装应用程序或进行额外的认证操作，这一机制使得攻击者能够绕过沙箱限制，进而读取受害者的私有对话数据1。
该漏洞的核心机制在于利用ChatGPT HTML预览功能与加载MathLive库的图形组件之间跨帧消息处理的缺陷1。攻击者借助YouTube oEmbed端点的JSONP回调特性成功绕过了内容安全策略（CSP）的限制1。在实际攻击场景中，攻击者可通过调用特定API接口枚举受害者的对话列表，获取未共享的私有对话内容，并将这些数据发送至外部服务器1。
OpenAI已于2026年10月26日确认修复此P1级漏洞，并确保两个前端版本均能拒绝此类恶意的跨帧消息1。
Security researchers disclosed a critical vulnerability in ChatGPT that allowed attackers to execute arbitrary JavaScript code and hijack user accounts simply by sending a standard shared link. The flaw, identified as a P1-level issue, enabled malicious actors to bypass sandbox restrictions and access private conversation data within a victim’s active login session 1. OpenAI confirmed on October 26, 2026, that the vulnerability had been patched, ensuring that both frontend versions now reject the specific cross-frame messages used in the attack 1.
The exploit leveraged a defect in the communication between ChatGPT’s HTML Preview feature and its Graph component, which utilizes the MathLive library for rendering. Attackers circumvented Content Security Policy (CSP) protections by exploiting the JSONP callback mechanism inherent in YouTube oEmbed endpoints 1. Once triggered, the vulnerability permitted the enumeration of a victim’s conversation list and the exfiltration of unshared private chat logs to external servers via calls to the GET /backend-api/conversations API 1. Victims required no interaction beyond opening the attacker-created malicious link, with no need for additional authentication or software installation 1.
评论
还没有评论，欢迎留下第一条。