A Shared ChatGPT Link Could Have Taken Over Your Account — OpenAI Fixed It in 8 Days
A ChatGPT link someone shares with you isn’t automatically safe to open
Security research firm Parameter found that an ordinary, public ChatGPT share link — the kind anyone can send you in an email, a support ticket, a chat message, or a social post — could be turned into a way to take over the account of whoever opened it. Per Parameter’s own technical writeup, “Do the Math: How We Took Over ChatGPT Accounts”, the victim didn’t need to install anything, log into a separate account, or join any workspace. They just had to open the link and click ChatGPT’s “Preview” button once.
Once that happened, code the attacker had written ran inside the victim’s own signed-in ChatGPT session — not the attacker’s. From there, it could call ChatGPT’s real backend the same way the victim’s own browser normally does: listing the victim’s recent conversations and pulling content out of a private chat that had never been shared with the attacker in the first place.
How a shared chat could run a stranger’s code
ChatGPT lets you share a conversation as a public link, and lets anyone who opens that link preview things rendered inside the chat, including math formulas. Parameter found that the internal messaging ChatGPT’s Preview used to draw those formulas — a component called MathLive — accepted instructions without checking who actually sent them. That let a malicious shared conversation smuggle in its own HTML.
On its own, smuggled HTML inside a sandboxed preview isn’t much of a threat. The second part of the chain made it one: Parameter chained that HTML injection with a separate gap in ChatGPT’s security rules, which blocked ordinary inline scripts but made an exception for YouTube. A quirk in how YouTube’s own embed system handled a certain parameter let the attacker’s HTML execute as code that the browser trusted as belonging to chatgpt.com itself — not to a stranger’s shared link. In Parameter’s early tests, that code ran within about 4.5 seconds of the Preview click.
You don’t need the technical plumbing to get the takeaway: a feature meant only to preview a math formula became a doorway into a signed-in account, because one step in a long chain trusted content it shouldn’t have.
A controlled test, not an active attack
This is important context for a business owner deciding how worried to be: nobody’s account was actually raided. Parameter reproduced the entire chain between its own test accounts, reported it to OpenAI, and OpenAI shipped a fix before Parameter published anything about it publicly. Independent reporting from The Hacker News confirms the same timeline:
- October 13, 2026 — Parameter first reproduces the underlying flaw.
- October 18 — Parameter submits a report to OpenAI’s bug bounty program.
- October 21 — OpenAI reproduces the issue and raises its internal severity rating from P2 to P1.
- October 26 — OpenAI ships a fix across both of ChatGPT’s web frontends. Eight days, start to finish.
OpenAI paid Parameter a $5,000 bounty for the report. Neither Parameter nor OpenAI has reported any real-world victim — this was disclosed, fixed, and written up in that order.
What this means if your business uses ChatGPT
There’s nothing to install, patch, or configure here — the fix is already live, and it applies automatically to every ChatGPT account. This isn’t a “your business got hacked” story; it’s an awareness story about a kind of risk that’s easy to overlook because it comes from inside a tool you already trust.
- No action is required right now. OpenAI’s fix closed this specific flaw on October 26, before it reached the public.
- Treat “someone shared a ChatGPT conversation with you” like any other link in an email — opening it and clicking around isn’t automatically risk-free just because the page says chatgpt.com.
- Remember that AI chat tools carry the same web-security risks as any other SaaS app your business logs into. A sharing or collaboration feature is still a feature with code behind it, and code can have bugs.
- If staff use ChatGPT’s share links to pass drafts, research, or customer answers around, this is a good moment to mention to your team that AI tools deserve the same basic caution as any other link-sharing software — not because this specific bug is still live, but because the next one might not get caught in eight days.
This is also the second time in two weeks that a “share this AI conversation” feature has turned out to carry more risk than it looked like — see our earlier report on how Meta’s Muse share links were publicly searchable. Different vendors, different bugs, same lesson: an AI chat tool’s “share” button is a piece of web infrastructure like any other, not a special safe zone.
The takeaway
A security firm found that one click on an ordinary ChatGPT share link could have handed an attacker a live, signed-in session with access to a victim’s private chats. OpenAI fixed it in eight days and paid a $5,000 bounty, and no real-world victims have been reported. The fix is already protecting every ChatGPT user today — the only thing worth carrying forward is the habit: a link that looks like someone sharing an AI conversation with you is not automatically safer than any other link in your inbox.
For more on what’s changing in AI for small business owners, see our knowledge base, and follow @AIWASai for the next story.
Stop leaving customers on read.
We answer your reviews, keep your social pages posting, and make your ads convert, so you get more customers while you run the business.
Get Started