Source: Parameter Security research disclosure published by Mukul Goyal, titled 'Do the Math: How We Took Over ChatGPT Accounts', disclosed October 31, 2026.
ChatGPT executes user-generated HTML code blocks inside isolated, cross-origin iframe sandboxes to prevent untrusted code from accessing top-level DOM properties or session cookies. However, when conversations incorporate mathematical graphs, ChatGPT loads the MathLive rendering library. The library initializes a global window message listener to receive configuration updates for its on-screen virtual keyboard but failed to validate the origin (event.origin) of incoming postMessage dispatches.
Parameter's exploit weaponized this omission by embedding an HTML Preview that dispatched two targeted messages to the top-level parent window. The first message updated MathLive's keyboard layout setting with a crafted HTML structure containing an execution payload, while the second instructed the library to render the virtual keyboard on screen. By injecting into the parent document context, the payload escaped the cross-origin iframe boundary and executed directly within the chatgpt.com origin.
To overcome ChatGPT's Content Security Policy, the researchers bypassed script restrictions by executing within the context of trusted Angular and UI gadgets already initialized in the page. The payload successfully issued authenticated GraphQL and REST calls to OpenAI backend endpoints, dumping conversation titles, reading unshared message threads, and demonstrating complete administrative control over the active session before OpenAI deployed hotfixes on October 26.