Security researchers at Parameter Labs have revealed a chain of flaws that turned a public ChatGPT share link into a full account takeover — no password, no install, no pasted code. In the most polished version of the attack, opening the link was the victim's only required action.
The write-up, titled "Do the Math: How We Took Over ChatGPT Accounts" and authored by Mukul Goyal, was published on October 28 and picked up by The Hacker News, with a write-up by Simon Willison following on October 31. OpenAI fixed the flaw on October 26, classified the report as P1, and paid a $5,000 bug bounty — a figure Willison called cheap for the severity.
How a share link became full access
ChatGPT renders model-authored HTML in a Preview pane inside a sandboxed iframe on a separate origin. That sandbox held: Preview code could not read the main page. But the same conversation could also contain an interactive graph, and that graph loads MathLive, a math-input library, into the main chatgpt.com page.
MathLive installs a listener for postMessage events that control its on-screen keyboard. The listener accepted messages from any sender with no origin check, and treated the keyboard-layout field as HTML, rendering it via innerHTML in the trusted page. From there, the researchers injected a new iframe through the srcdoc attribute with no sandbox attribute, giving it the chatgpt.com identity.
ChatGPT's Content Security Policy blocked inline scripts but allowed YouTube, and YouTube's oEmbed endpoint reflected attacker JavaScript back in a JSONP-style response — the final step to full script execution on the chatgpt.com origin. Code running as chatgpt.com could call ChatGPT's authenticated APIs as the victim.
What the researchers could do — and how fast it was fixed
In their proof of concept, the researchers enumerated the victim's conversation list, read a private conversation that had never been shared, and exfiltrated data to an external server via an image request. In one account, the alert fired after roughly 4.5 seconds with no interaction at all, and the result was reproduced three times, including after a hard refresh.
The timeline moved fast. Parameter confirmed the attack path on October 13, built cross-account share delivery by October 16, and submitted the report on October 18, adding a private-conversation proof the next day. OpenAI reproduced it on October 21, raised the severity from P2 to P1, shipped a frontend fix on October 22, tightened the Content Security Policy on October 23, and confirmed both web frontends fixed on October 26 — eight days from report to fix.
No CVE was assigned, no evidence of exploitation in the wild has been reported, and OpenAI has published no advisory of its own; the fix is documented only in the researchers' timeline. MathLive 0.111.2 added same-origin validation for postMessage, closing the vector for every app embedding the library, not just ChatGPT. The patch was server-side, so users need do nothing.
Why the $5,000 bounty stung
Willison and Hacker News commenters alike called $5,000 low for a one-click, full-account takeover, with the worry that researchers may not bother reporting the next one. The deeper lesson, Parameter argues, is about seams: no single component failed. The sandbox held, the policy was in place, and MathLive is a legitimate library — the failure lived in how they combined.
For anyone shipping artifact previews, chart widgets or connector interfaces, the checklist is concrete: authenticate every cross-frame message by origin and source, keep dynamically created frames sandboxed along the whole rendering path, and treat any Content Security Policy allowlist entry for a third-party script endpoint as a potential execution path.



