A booby-trapped ChatGPT share link could have exposed users' private chats, researchers say
OpenAI fixed the flaw Oct. 26, and users don't need to do anything.
Opening a shared ChatGPT conversation could have let an attacker take over the viewer's account and read their private chats, security firm Parameter said in an Oct. 28 report. OpenAI fixed the flaw Oct. 26, and users don't need to do anything, the researchers said.
ChatGPT lets people share a chat through a public link. Parameter built one that ran the attacker's code in the signed-in session of whoever opened it, in some versions after a single click of ChatGPT's Preview button.
How it worked
Preview turns web code in a chat into a small working web page. ChatGPT runs that page in a walled-off "sandbox" so it can't reach the main site. But the shared chat also drew a graph, which loaded MathLive, a code library for typing math, into the main ChatGPT page. MathLive accepted on-screen keyboard messages from any sender, and those messages could carry raw web code. That let the sandboxed page slip code into ChatGPT itself, Parameter said.
The code could then list a victim's chats and copy a private one to the researchers' server. Parameter said it tested only on its own accounts, and no real-world attacks were reported.
OpenAI rated the flaw in its most serious category and paid a $5,000 bug bounty, a reward for outsiders who find security holes, The Hacker News reported.
Source: Parameter, Oct. 28, 2026 and The Hacker News, Oct. 31, 2026.