Do the Math: How We Took Over ChatGPT Accounts
Parameter researchers describe a now-fixed ChatGPT exploit where opening a malicious shared conversation could run attacker code in a signed-in user’s session and expose private conversations.
The chain combined an unauthenticated cross-frame MathLive message, HTML rendered into an unsandboxed frame, and a permitted YouTube JSONP endpoint to bypass CSP. The researchers say OpenAI fixed it October 26. HN commenters questioned the reported $5,000 bounty and discussed the risks of interacting features.