31 October 2026
Researchers show ChatGPT previews could be hijacked to steal signed-in user data
First reported
Simon Willison ran this on .
- Parameter Labs, a security research group, disclosed a multi-step attack against ChatGPT's HTML preview feature.
- The attack chained two weaknesses, one in a math-display library and one in how YouTube embeds are permitted to load.
- Together they could run hostile code in a signed-in user's session and pull out private data.
How it was covered
Simon WillisonDaily notes and links
The attack chain, the MathLive virtual keyboard flaw, the YouTube oEmbed CSP bypass, data exfiltration, OpenAI's fix, and the bounty amount