A booby-trapped ChatGPT share link could have read your private chats. OpenAI has fixed it.

Researchers at the security firm Parameter showed that a public ChatGPT share link, the kind you send to show someone a conversation, could be set up to run the attacker's code inside the signed-in session of whoever opened it (Parameter). In their test, that code listed the victim's recent chats, opened a private one that had never been shared, and sent part of it to the researchers' server. Because the code ran as chatgpt.com itself, the researchers wrote: "Essentially, we could perform any action on behalf of the victim's ChatGPT account." The trick chained several features together: an interactive graph in the chat loaded a math keyboard tool that accepted instructions from ChatGPT's sandboxed HTML Preview without checking where they came from, which let the Preview slip code into the main page. In the most reliable version, the victim just had to open the link and click Preview once, and a later version opened Preview by itself. Parameter reported the flaw on Oct 18. OpenAI raised its severity to P1 and fixed it on Oct 26, and paid a $5,000 bounty. Parameter published its writeup on Oct 28, and The Hacker News covered it today. Parameter ran the test between its own accounts and doesn't report any real-world use.

What this means for you: there's nothing to install or reset, because the fix was on OpenAI's side. It's a good reminder that a shared AI chat is a web page like any other, so a link from someone you don't know deserves the same caution as any other link.