Parameter Labs details how a ChatGPT HTML preview exploit could take over accounts
AIParameter Labs describes an exploit against ChatGPT's HTML preview sandbox that chained a MathLive virtual keyboard message handler with a YouTube oEmbed endpoint to run JavaScript. The attack could fetch private user data and exfiltrate it through an image reference when a signed-in user visited a malicious share URL. OpenAI fixed the hole and paid a $5,000 bug bounty, which Simon Willison calls cheap for a vulnerability this serious.
AIWhy it matters.
The write-up traces a multi-step sandbox escape through a third-party library and a CSP allowlist, showing how one weak link can expose user data.
Source details
Simon Willison · simonwillison.netPublished · added here
Original title: How We Took Over ChatGPT Accounts