howitworksEk ChatGPT link se saari private chats leak? Poora hack step by step
Ek public ChatGPT share link, logged-in user ke kholte hi, uske session mein attacker ka JavaScript chala sakta tha. Chain: HTML Preview ek sandboxed iframe (alag origin) mein chalta tha, par graph ke liye load hui MathLive library ka postMessage listener originValidator "none" par tha. Preview ne keyboard layout ke markup mein HTML bheja, MathLive ne use innerHTML se page mein lagaya, usme ek srcdoc iframe tha bina sandbox attribute ke, jise chatgpt.com origin mila. CSP mein www.youtube.com allowed tha, aur YouTube oEmbed ka JSONP callback attacker ka code chala deta tha. Code ne page se access token liya aur /backend-api/conversations se private chats nikaali. Parameter ne 18 Oct ko report kiya, OpenAI ne 26 Oct tak fix kiya (P1, $5,000 bounty). Kisi asli user par hamle ka saboot nahi mila.
Links
Aaj ye karo
- Anjaan ya unexpected ChatGPT share link incognito / private window mein kholo, jahan aap logged in nahi ho.
- ChatGPT: Settings > Data controls > Shared links > Manage, aur purane shared links delete karo.
- Chat mein kabhi password, API key ya token paste kiya hai toh use turant badlo / revoke karo.
- Developers: postMessage listener mein event.origin aur event.source dono check karo, aur untrusted HTML ko kabhi innerHTML se mat lagao.
Sources
Parameter, The Hacker News, Ostorlab, DEV Community, OpenAI Help; 1 Nov 2026 ko check kiya.