Researchers at Parameter Labs discovered a security flaw in ChatGPT’s HTML preview feature that enables attackers to hijack user accounts. The vulnerability was triggered when a victim opened a malicious share link sent by an attacker. The researchers confirmed the flaw worked by taking over a ChatGPT account without the victim installing apps or using developer tools.
The attack chain used the MathLive library to inject HTML into the main page via a postMessage API. This allowed the attacker to load a YouTube oEmbed endpoint to execute arbitrary JavaScript. The executed code could enumerate the victim’s conversation list and access private data. OpenAI patched the vulnerability on October 26, 2026, and paid a $5,000 bug bounty.
