A public ChatGPT conversation could carry code that read private conversations from another signed-in user's session. Parameter Labs' October 28 report says OpenAI fixed the flaw on October 26.

The surprising part: the HTML Preview's isolation worked. A separate math-keyboard component in the main page accepted messages from the wrong sender, letting untrusted content cross the boundary. Several ordinary features combined into the attack.

Mukul Goyal's account describes a controlled test between accounts, including private-chat data leaving the session. It is a disclosed, reportedly fixed vulnerability, not evidence of ongoing exploitation. The useful design lesson: an isolated frame still needs strict checks on every component allowed to receive its messages.