A shared ChatGPT link was enough to read private chats: what to check before opening one
Researchers showed that opening a public ChatGPT link could run someone else's code inside the account of the person opening it and read their conversations. According to their report, OpenAI has already fixed the flaw and paid them $5,000, but the case holds lessons for users and businesses.
Photo: Brendan Cronin on Unsplash
Sharing a ChatGPT conversation with a colleague is almost automatic: you copy the link, paste it into the work chat, and you're done. That habit was the entry point for a flaw that, according to a report by the research firm Parameter Labs published on October 28, 2026, let an attacker craft a public ChatGPT link capable of running their own code inside the account of whoever opened it and, from there, reading that person's private chats.
The most troubling detail is who was at risk. It wasn't the person sharing the link but the person opening it. According to the report, written by researcher Mukul Goyal, the victim only had to open the link. In the most reliable version of the attack, they had to click the "Preview" button once, although the researchers later managed to make that preview open on its own.
A crack in the previews
The problem lay in the ChatGPT feature that displays small interactive web apps inside a conversation, previously known as Canvas. These apps appear in an isolated frame, a kind of sealed box that keeps their content from touching the rest of the page. According to Parameter Labs, that box worked as intended. The flaw was in how ChatGPT's main page handled the messages it received. A third-party tool for displaying math formulas, called MathLive, accepted those messages without checking where they came from. Combined with other oversights, this let the attacker slip in instructions under the identity of chatgpt.com itself.
With that access, the researchers explain, their code used the temporary key that keeps the user's session open to list their recent conversations, read a private chat that had never been shared, and send an excerpt to a server they controlled. They note that in theory the code could do anything on the victim's behalf, but they only tested reading chats. They did not demonstrate account changes or access to files, and there is no public evidence that anyone exploited the flaw outside that test.
According to the report's timeline, Parameter Labs reported the flaw to OpenAI's bug bounty program on October 18. The company reproduced it on the 21st, raised its severity to P1, and confirmed on the 26th that it had been fixed. That was eight days. Because the fix was made on OpenAI's servers, users didn't have to update anything. The bounty was $5,000, according to the researchers and developer Simon Willison, who publicized the case on October 31 and said the amount was "a bit cheap" for something so serious. That program, run on the Bugcrowd platform, has paid up to $100,000 for exceptional critical findings since March 2025, according to BleepingComputer. OpenAI has not made a public statement about the case or the payout.
What to check
This isn't the first time ChatGPT's shared links have made headlines. In 2025, conversations turned up indexed in search engines, and OpenAI removed the option to make them "discoverable." Its chief security officer, Dane Stuckey, acknowledged at the time that the feature introduced "too many opportunities" to share something unintentionally. That case was about what people publish. This one shows that opening a link can also have consequences.
No source mentions affected users in Latin America, and neither OpenAI nor security agencies in the region have published guidance on this case. Even so, a few common-sense precautions, which are recommendations from our newsroom, reduce the risk. It's wise to open shared links only if they come from people you know and to log out when you're not using the tool. It also helps to delete links you no longer need and to avoid pasting passwords, customer data, or company code into a chat. In organizations, it's worth setting clear rules about what can be shared with an AI. One important detail: two-step verification would not have stopped this attack, because the code ran inside a session that was already open.
It remains to be seen whether OpenAI will independently confirm the dates and the amount, whether a formal security advisory will be issued, and whether other chatbots that also let users share conversations will review similar features. In the meantime, an AI link deserves the same suspicion as any other link that arrives without context.