Parameter Labs Hijacked ChatGPT Accounts Through a Sandboxed HTML Preview
We Took over ChatGPT Accounts
Parameter Labs chained several small flaws into full account takeover on ChatGPT. MathLive's virtual keyboard accepted postMessage() calls from any sandboxed iframe and injected HTML via innerHTML, letting untrusted code create an unsandboxed iframe. CSP allowed scripts from YouTube, whose oEmbed endpoint reflected JavaScript back for execution. With script running, private data was exfiltrated through an image reference — so a shared chat link could steal a signed-in visitor's data. OpenAI patched it and paid $5,000.
OpenAI closed the hole and paid out a $5,000 bug bounty, which I think was a bit cheap for a vulnerability this serious.