A shared ChatGPT link could read private chats: two rules for businesses

A shared ChatGPT link could run code in the session of the person who opened it and read their private chats. OpenAI has fixed the vulnerability, but two rules remain for companies that use AI.

Sharing a conversation seems harmless: you just send someone a link. However, researchers at Parameter have shown that a specially crafted sharing link could execute code right within the session of the person who opened it—not on some obscure website, but on chatgpt.com, under that person’s identity.

In the test, the attacker’s code listed the victim’s recent conversations, opened a private chat that had never been shared, and sent part of its content to a server controlled by the researchers. In the most reliable version, the victim only had to open the link and click “Preview” once; in a later version, the page automatically triggered “Preview.” No installation or authentication was required.

At the core were two vulnerabilities: a mechanism that accepted raw HTML messages from any sender and a framework that inherited the trust of the main page, treated as if it were part of chatgpt.com. The researchers say they could “perform any action on behalf of the victim’s ChatGPT account.”

OpenAI classified the issue as P1, the most severe category, and fixed it on the server side. You don't need to install anything, and the research was controlled—there were no known real-world attacks. What matters is what remains true after the vulnerability has been patched.

What you need to do:

  • Treat a link to a shared AI chat like any other unsolicited link: don’t open it and don’t click “Preview” if it comes from a source you don’t trust.
  • Set a simple rule for the team: what NOT to include in AI chats—customer data, passwords, contracts. It was precisely the content of private conversations that could have been leaked here.
  • Decide who in the company is allowed to link work accounts (email, files) to AI tools.

Whatever you type into an AI assistant is still company data. The breach has been fixed, but the lesson remains: an AI conversation is not a guaranteed private space, and a link to an AI chat is active content, not just text.

This article was generated with AI assistance.

Request a quote

← All news