Researchers at security firm Parameter showed that a malicious ChatGPT share link could run code in the signed-in session of anyone who opened it, list their recent conversations and read private ones, The Hacker News reports. The attack combined the HTML Preview feature with a maths library that accepted messages from any frame. OpenAI received the report on 18 October, rated it in its most serious category and fixed it by 26 October. There is no sign of real-world exploitation.

What it means: Many employees now use ChatGPT for quotes, contracts, code and customer details, so an account's chat history is effectively a file of company information. This weakness was dangerous because it required nothing unusual: the link looked like an ordinary shared conversation and could arrive by email, support ticket or social post. The good news is that researchers found it through the bug bounty programme and the fix came within eight days. The broader lesson is that AI apps are quickly adding features that execute content, such as web page previews and interactive charts, and each of them opens a new path for attack. For Greek businesses, this strengthens the case for centrally managed company accounts instead of personal ones, a policy on what may be put into AI tools and regular clearing of history that is no longer needed. Shared chat links from unknown senders also deserve the same suspicion as an unknown attachment, because the next weakness may not be found by researchers first.

Read the original on The Hacker News

Related service: Custom software

Related stories