Parameter Labs reveals the attack chain for taking over a ChatGPT account
Simon Willison reports on a ChatGPT sandbox escape vulnerability chain disclosed by Parameter Labs. Untrusted code within an HTML preview iframe can trigger MathLive's virtual keyboard via postMessage, injecting innerHTML content. This is then reflected and executed as JavaScript through a YouTube oEmbed endpoint permitted by Content Security Policy (CSP), potentially allowing attackers to steal private data from logged-in users or perform actions on their behalf.
Return to original text
A short introduction is retained here; the full text from the source website is not reposted. For specific wording and subsequent updates, please refer to the original article.
Open AIHOT original