How We Took Over ChatGPT Accounts
Reported by Hacker News and Simon Willison2 sources ↓
Parameter Labs researchers reported that a malicious ChatGPT share link could run code in a signed-in user's session, exposing private conversations and allowing actions through the user's account. OpenAI fixed the issue on October 26, 2026.
Key points
- Opening a malicious public share link was enough to run attacker-controlled code in the recipient's signed-in session
- Researchers said the code could retrieve private conversations and call ChatGPT APIs as the user
- OpenAI fixed the issue on October 26, 2026, and classified the report as P1
- Simon Willison reports that OpenAI paid a $5,000 bug bounty
Sources
Details
Parameter Labs described an attack in which a malicious ChatGPT share link could execute JavaScript in the session of a signed-in user who opened it. The researchers said they could enumerate conversations, retrieve content from a private conversation, send it to an external server and call authenticated ChatGPT APIs as the user. The recipient did not need to install an app or take any other action beyond opening the link.
The attack began with code in an HTML Preview sending messages to a MathLive listener in the main ChatGPT page. The listener accepted messages without checking their origin and rendered supplied HTML; the researchers used that path to add an unsandboxed iframe. They then used a YouTube oEmbed response to run JavaScript despite ChatGPT's content security policy blocking inline scripts.
Parameter Labs said the original delivery method did not work through a public share because the graph needed for the attack was absent from the shared conversation. The researchers described using a native HTML attachment in the Preview as another route. OpenAI fixed the issue on October 26, 2026, and classified the report as P1; Simon Willison reports that the company paid a $5,000 bounty.
Written from coverage as of