Summary & Key Takeaways ​

• A complex exploit allowed account takeover and data exfiltration from ChatGPT.
• The attack leveraged vulnerabilities in HTML preview sandboxing.
• It involved postMessage(), innerHTML injection, and YouTube oEmbed reflection.
• JavaScript execution was achieved despite strict Content Security Policy rules.
• The exploit could fetch private user data and act on their behalf.
• OpenAI patched the vulnerability and issued a $5,000 bug bounty.
• The article details the intricate steps of the attack chain.

Our Commentary ​

This is wild. A truly intricate attack chain, jumping from sandboxed iframes to innerHTML to YouTube oEmbeds for script execution. It's a masterclass in finding obscure interactions. I'm genuinely surprised by the $5,000 bounty for something this severe; it feels incredibly low for a full ChatGPT account takeover. It makes me wonder about the true value placed on user data.