OpenAI patched ChatGPT share-link flaw that read private chats
November 2: GBHackers adds a $5,000 bounty, October 21 reproduction and P2-to-P1 upgrade, fixes on both web frontends, and no reported real-world attacks.
A fixed P1 ChatGPT bug let a public share link run script in a signed-in session and read private chats.
Parameter researchers showed that a public ChatGPT share could run attacker JavaScript inside the signed-in chatgpt.com session of anyone who opened the link. A Preview frame and a MathLive graph listener turned untrusted postMessage content into HTML, then an unsandboxed srcdoc iframe and a YouTube JSONP callback bypassed CSP. The script could list conversations, read an unshared private chat, and call authenticated APIs, including the session token and account ID, according to the write-up. OpenAI reproduced the issue on October 21, raised it from P2 to P1, paid a $5,000 bounty, and by October 26 both web frontends rejected the cross-frame MathLive messages. The write-up does not say this flaw was used in attacks.
- Opening the shared link was the only victim action; payload ran after about 4.5 seconds.
- MathLive accepted update-setting messages from any origin and set innerHTML.
- CSP allowed YouTube scripts, abused as an oEmbed JSONP callback.
- Injected code called /backend-api/conversations and exfiltrated part of a private chat.
- OpenAI reproduced it October 21, raised it P2 to P1, and fixed both web frontends October 26.
- OpenAI paid a $5,000 bounty; in-the-wild use is not reported.
Timeline
dated events in the reporting · oldest first- Oct 21, 2026OpenAI reproduced the bug and raised it from P2 to P1 · GBHackers
- Oct 26, 2026Both ChatGPT web frontends blocked cross-frame MathLive messages · GBHackers
- Oct 31, 2026Researchers published the share-link account-takeover write-up · Hacker News
- Nov 2, 2026Report notes $5,000 bounty and no known attacks · GBHackers
What each outlet adds
beyond the shared facts- Hacker NewsUnsandboxed srcdoc iframe and confirmed private-chat read plus API calls
- Simon WillisonNames Parameter Labs and ChatGPT’s HTML-preview artifacts, formerly Canvas
- GBHackers$5,000 bounty, Oct 21 repro, P2-to-P1, both frontends fixed
- Vendors
- OpenAI
- Threat actors
- Parameter Labs
- Malware
- NetSupport RAT
- Organizations
- Parameter
Coverage
oldest first · one row per article · UTCChatGPT share links executed JavaScript in signed-in sessions
Researchers read a private chat and sent it off-site; OpenAI fixed the P1 issue on October 26.
Parameter Labs finds exploit to take over ChatGPT accounts
Parameter Labs discovered a complex attack chain exploiting ChatGPT's HTML previews via postMessage-triggered MathLive virtual keyboard.
ChatGPT share-link chain let attackers run script in signed-in sessions
OpenAI rated the issue P1, paid a $5,000 bounty, and blocked the cross-frame MathLive messages on both web frontends by October 26.