ChatGPT's Lockdown Mode had a hole, and it was plugged in September

Lockdown Mode is the switch OpenAI added for people who want ChatGPT kept on a short leash. It reached personal accounts in June, under Settings and then Security. On Thursday 5 November Check Point Research published a way around it, found by researcher Alexey Bukhteyev, and the good news comes first: OpenAI fixed it on 18 September, and Check Point confirmed the trick no longer works.

The weak spot was the HTML preview that ChatGPT has shown under code blocks since February. That preview runs in a sandbox, and ChatGPT tried to stop it opening new windows. But the sandbox still allowed popups, and a page could slip in one more frame and use that frame's own window to open a popup that escaped. From there an attacker could show a window dressed up as a ChatGPT update that delivered malware, or sign the user out and catch the password on a fake login page. In the older ChatGPT Classic desktop app it didn't even need a click: having the conversation was enough to send the whole chat history and uploaded files away quietly.

Nothing suggests anyone used it before the fix, and there's nothing you need to install. The lesson outlasts the bug, though. A chat someone shares with you is a web page someone else wrote. ChatGPT doesn't update itself through a popup inside a conversation, and its sign-in page sits on an OpenAI address, so a prompt to download something or log in again from inside a chat deserves a hard look at the address bar first. If you still use the Classic desktop app, keep it current.

Tendvane's Privacy & accounts check is a good place to review how your important logins are protected, so one phished password doesn't open everything.