“Opening the shared link was the only required action”: ChatGPT privacy flaw is fixed

“Opening the shared link was the only required action”: ChatGPT privacy flaw is fixed

A malicious ChatGPT shared-conversation link could have let an attacker read private chats from a signed-in user’s browser session, security researchers say. OpenAI fixed the issue on October 26, according to the research team that reported it.

The researchers at Parameter showed that specially prepared content in a shared conversation could run code with the permissions of the recipient’s ChatGPT session. In a controlled proof of concept, it listed conversations, retrieved one that had never been shared with the attacker and sent a small selected piece of its data to a researcher-controlled server.

In its October 28 technical report, Parameter says it reported the issue through OpenAI’s bug bounty program on October 18. The team says OpenAI reproduced the problem, raised its severity to P1 and confirmed that both versions of the affected web frontend rejected the reported messages by October 26.

How a shared conversation crossed the boundary

ChatGPT’s HTML Preview runs code in a restricted frame separate from the main page. The researchers say that isolation worked when tested directly, but an interactive graph loaded a math component into the main ChatGPT page that accepted messages from the preview without adequately checking their origin.

The team combined that message path with other browser behavior to run code as chatgpt.com. That made the difference between merely altering the shared page and using the recipient’s authenticated session to request data from ChatGPT’s conversation interface.

In the most reliable version of the researchers’ test, the recipient opened the shared conversation and clicked its Preview button. A later version opened the preview automatically; Parameter says it reproduced that path three times on a separate account. “Opening the shared link was the only required action,” the researchers wrote about that version. The proof used accounts controlled by the researchers.

What users should take from the fix

The Cloud Security Alliance’s analysis says the fix was deployed on the service side and users do not need to take action for this particular issue. Parameter’s report says OpenAI resolved it; neither source establishes that a reader’s account was compromised.

The finding concerns a malicious shared link viewed in a signed-in web session. It is different from intentionally sharing a conversation with someone: in the researchers’ demonstration, the exposed private conversation had not been shared with the attacker.

Unsolicited links still deserve care even when their address belongs to a familiar service. The Cloud Security Alliance says no end-user action is needed to receive the server-side fix for this particular issue.

Leave a Comment

Your email address will not be published. Required fields are marked *