AivexaNewsSearch
AI news for builders and product teamsChecked every hour

How We Took Over ChatGPT Accounts

Collected Nov 1, 2026

Parameter Labs published an account of an exploit chain it found in ChatGPT's implementation of what OpenAI formerly called Canvas and now appears to call HTML previews, the feature that renders interactive custom HTML apps inside a sandboxed iframe. The core requirement for that feature is tight isolation: untrusted HTML must not be able to reach OpenAI's APIs or touch a user's private data. Parameter's chain, as described by Simon Willison, broke that isolation in several stages and ended with a crafted share URL that could exfiltrate another signed-in user's data or act on their behalf. OpenAI has closed the hole and paid a $5,000 bug bounty.

The first link was MathLive, a JavaScript library OpenAI uses for graph support. MathLive ships a virtual keyboard, and the parent page can be told to open it via postMessage() calls coming from the sandboxed iframe. The code that receives those messages writes into the parent page using innerHTML. Nothing prevented frames other than the intended one from sending that message, so an HTML preview frame could reach the handler too. Because innerHTML does not execute script tags directly, the practical outcome was not immediate code execution but markup injection: an attacker could plant a fresh iframe in the parent page without the sandbox restrictions applied to the original preview.

That still left the problem of running JavaScript. OpenAI's Content Security Policy blocked inline scripts entirely and permitted script loading only from a short list of domains. One of those domains was YouTube, and the YouTube oEmbed endpoint could be coaxed into reflecting JavaScript back in a form the browser would execute. Combining the injected unsandboxed iframe with that reflected script gave the attackers JavaScript execution in the context of ChatGPT itself. From there, the writeup describes fetching private user data through APIs and exfiltrating it by embedding it in an image reference.

The payload had a delivery mechanism that fits neatly into normal product behavior: a ChatGPT share URL. A signed-in user who opened the link could have their private data pulled out or could be made to perform actions under their own session. Willison notes the $5,000 payout and argues it looks cheap for a vulnerability of this severity.

For developers building similar sandboxed rendering features, the chain is a useful catalogue of failure modes. Message-based bridges between frames need origin and identity checks, and any handler that writes untrusted strings into the DOM should be treated as a sink, even when it is only innerHTML. CSP allowlists deserve scrutiny for endpoints that reflect content, since a trusted domain that mirrors user-influenced input can serve as a script transport. Each individual step here looks like a minor weakness; the impact came from composing them.

Why it matters: anyone shipping iframe-sandboxed user content, especially with postMessage bridges to a privileged parent page, inherits versions of this risk. For product teams, the fact that a share link alone could reach another user's data is a reminder that shared-content surfaces are attack delivery paths, and that bug bounty pricing signals how seriously such chains are valued. Expect, as an inference rather than a stated fact, tighter review of third-party widget libraries and of allowlisted script sources in similar platforms.

Read at Simon Willison

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

How We Took Over ChatGPT Accounts Parameter Labs describe an exploit they found against ChatGPT's implementation of the artifacts pattern (OpenAI used to call this "Canvas" but it looks like the latest revision is called HTML previews ), where interactive custom HTML apps are displayed in a sandboxed iframe. It's important that HTML in these stays tightly sandboxed to avoid untrusted code interacting with OpenAI APIs and accessing private data. Parameter found quite a complex attack chain. OpenAI use the MathLive J