A critical cybersecurity vulnerability in the ChatGPT web platform allowed cybercriminals to hijack the active session of any registered user and steal confidential conversations by convincing the victim to open a link with an apparently harmless shared conversation, after which malicious code took control in the background, accessed private history, and sent data to an external server. The breach stemmed from a communication failure between the HTML code preview and the MathLive math library: the preview operates within an isolated sandbox, but MathLive kept a listener open for messages to activate its virtual keyboard without properly verifying the sender, allowing the isolated environment to send instructions to the main site and inject malicious HTML code into the AI interface. Researchers from the firm Parameter created a floating frame (iframe) that inherited ChatGPT's identity and full permissions and used a legitimate YouTube connection point on the platform's trusted sources list to bypass browser shields and chatgpt.com's security policies. In the final demonstration, the victim only had to click the public conversation link, and in just 4.5 seconds the script was scanning OpenAI's internal APIs to extract private chats through an image request. The breach was reported on October 18 by researcher Mukul Goyal on behalf of Parameter; OpenAI categorized it as priority P1, fixed the flaw on its servers on October 26, and rewarded the team with $5,000, and the vulnerability has already been resolved without users having to do anything.