OpenAI has addressed a significant vulnerability in ChatGPT that allowed attackers to expose private conversations by tricking signed-in users into participating in a malicious shared conversation This article explores vulnerability chatgpt. . Researcher Mukul Goyal revealed the issue on October 28, and OpenAI resolved it on October 26, classifying it as P1.
The exploit combined public conversation sharing, HTML Preview, the MathLive mathematics library, and a content security policy bypass. OpenAI Corrects High-Severity ChatGPT Vulnerability This prevented previewed JavaScript from directly accessing the surrounding ChatGPT page.
However, an interactive graph loaded MathLive onto the main page, setting up a listener for virtual keyboard messages. Although scripts inserted through innerHTML remained inactive, the injected markup created an unsecured iframe using srcdoc. Mukul Goyal exploited a JSONP callback vulnerability in YouTube’s oEmbed endpoint, allowing him to inject malicious JavaScript into the victim’s session.
He enumerated recent conversations, retrieved messages from an unrelated private conversation, and transmitted a small payload to an external server via an image request.
OpenAI reproduced the execution and private-conversation disclosure, then updated two web frontend versions to reject the reported cross-frame messages. Effective defenses necessitate authenticating message origins and source windows, excluding raw HTML from cross-frame settings, and maintaining sandbox restrictions around untrusted documents.
