A shared ChatGPT link could run attacker code in your session and read private chats. OpenAI fixed it on 26 October
Security firm Parameter published a write-up on 28 Oct 2026 of a flaw chain in ChatGPT's shared-chat pages. Opening a malicious shared link, with at most one click on Preview, could run an attacker's code as chatgpt.com and read a private conversation. Parameter says OpenAI fixed it on 26 Oct and asks nothing of users. What happened and a simple habit worth keeping.
How urgent
Good to know
Who is affected
People who open ChatGPT shared-chat links while signed in. Parameter says OpenAI fixed the issue on 26 October 2026 and describes no action users must take
What to do
Nothing is required according to the researchers. As a habit, open shared ChatGPT links only from people you trust, and review the shared links you have created under Settings, Data controls, Shared links, deleting any you no longer need
Published 3 Nov 2026 · Sources last checked 3 Nov 2026No. 43 / 87
What happened
On 28 October 2026 a researcher at the security firm Parameter published a write-up of a flaw chain in ChatGPT's shared-chat pages. According to the write-up, an attacker could build a conversation, publish it with ChatGPT's normal Share feature, and send the link. When a signed-in person opened it, the page could run the attacker's code as chatgpt.com, inside the victim's session.
How it worked, in plain terms
The write-up describes three weaknesses that, chained together, crossed a boundary that was meant to hold:
- A shared chat with an interactive graph loaded a math library called MathLive into the main ChatGPT page. That left a listener running there.
- A preview of an HTML block inside the same chat sent that listener messages. The listener accepted them without checking who sent them and treated raw HTML as a keyboard layout, which put the attacker's HTML into the page.
- That HTML opened a frame that was not sandboxed. ChatGPT's page rules blocked inline scripts but allowed YouTube, and the researcher used a YouTube feature to get script running.
What the victim had to do: in the most stable test, click the normal Preview button once. In a later version, the HTML block was set to preview automatically, so opening the link was enough.
What the attacker could reach
According to the write-up, the code could list the victim's recent conversations and read a private conversation that had never been shared with the attacker, sending a small piece of it to an outside server. The researcher argues it could also call ChatGPT's signed-in APIs as the victim. That last point is the researcher's inference, not something the write-up shows working.
Timeline and fix
- 18 Oct: reported to OpenAI's bug bounty program.
- 21 Oct: OpenAI reproduced it and raised the severity from P2 to P1.
- 22 and 23 Oct: one web frontend was fixed first and changes to the page rules blocked part of the chain, while a second frontend was still being updated.
- 26 Oct: OpenAI confirmed both frontends rejected the messages, marked it resolved and paid a $5,000 bounty in total.
The write-up asks nothing of users. The fix is on OpenAI's side.
A habit worth keeping
Our advice, not the researchers': a shared AI-chat link is a web page someone else built, so treat it like any link from a stranger.
- Open shared ChatGPT links only from people you trust.
- Review the shared links you have created under Settings, Data controls, Shared links. Anyone who has a link can read that conversation and pass it on, so delete the ones you no longer need.
- Keep sensitive material out of chats you plan to share.
What we could not confirm
We read the researchers' own write-up. We could not read an OpenAI statement. Other sites say a new MathLive version, 0.111.2 on 31 October, changed the library's default so it checks where messages come from, but the write-up does not say the library changed, and we did not verify it. The Hacker News article on this story would not open for us. We do not know whether anyone used the flaw in real attacks before the fix.
Comments
No account needed. Add a nickname if you like, or leave it empty to comment anonymously. Comments are reviewed before they appear. Links are not allowed.
We store your comment, your nickname and a scrambled form of your connection that we use only to limit spam. We do not store your IP address or set cookies. Cloudflare Turnstile checks that you are not a bot.
Sources and verification
We use primary sources first and say when we rely on a news report. Dates are the day we last checked each source.
Spotted a mistake? Email [email protected] and we will correct it and note the change.