Researchers hijack accounts via ChatGPT share links; OpenAI fixes the flaw

We Took over ChatGPT Accounts

A research team found that a malicious public ChatGPT share link could run JavaScript inside a visitor's logged-in session, read private conversations and send them out. The chain used MathLive in chart loading, HTML Preview and YouTube JSONP; the auto-preview version fired on opening the link alone, and the team reproduced it three times on another account. OpenAI confirmed on October 26 that both web frontends are fixed, rated the bug P1 and paid $5,000 in total bounties.

Why it matters: The team traced the full path of a share link across sandboxes, showing that cross-window message checks and untrusted HTML isolation must hold through combined component rendering.