A crafted link could exploit the ChatGPT share link flaw and steal private data. OpenAI has now closed that hole in its chat service. The flaw sat in the HTML preview feature, where untrusted code clearly needs firm sandbox controls. Pivot News reported the chain on November 1.
Parameter Labs found the chain, while OpenAI paid a 5000 dollar bug bounty. Developer Simon Willison summarised the research on October 31. Simon Willison called that amount a bit cheap for a flaw this serious. The payment matters less than the ChatGPT share link flaw fix, however, because the attack needed only a link visit.
The chain began with MathLive, a maths library OpenAI uses for graph support. That library ships its own virtual keyboard, for example, for on screen maths input. The keyboard could wake up in the main page through a message from a sandboxed frame. The code that received that message built page HTML directly from the message content.
Nothing, however, stopped other sandboxed frames from sending that message. HTML preview frames counted among those frames, for example, and that widened the opening. The researchers used the ChatGPT share link flaw to inject a fresh frame with no sandbox limits. Direct script tags do not run through this route, as a result, so the team needed one more step.

OpenAI content rules, however, blocked inline scripts and allowed scripts from a short domain list. YouTube sat on that list, but its oEmbed endpoint could reflect JavaScript back in runnable form. With code running on ChatGPT itself, fetching private data became trivial, as a result. After that, the team could send that data out through an image reference.
Share links made the ChatGPT share link flaw dangerous for ordinary users. Victims needed no prompt, no file download, and no strange click beyond opening a link. A signed in user who opened a crafted link could lose private data, however. That user could also see actions taken in their name without consent.
OpenAI documents the ChatGPT share link flaw preview feature as broadly available across plans and devices. Availability varies by plan, device, and workspace settings in practice. Admin controls, at the same time, govern code execution and network access for workspaces. Teams can tighten the feature further through those controls, meanwhile, while they wait for vendor patches.
The preview feature grew out of Canvas, an older OpenAI workspace tool. It now covers HTML pages, React components, SVG images, Mermaid diagrams, and Vega charts. Each format runs inside the same sandbox promise. Each format must prove that promise against creative attacks in turn.
Users should treat shared links with the same care they give email attachments. A link from an unknown sender deserves a pause before opening. A link that arrives out of context deserves a second look as well. Account holders can sign out on shared devices, which removes the signed in state the attack required.

Bug bounties shape security work in public, for better and for worse. A 5000 dollar payout for the ChatGPT share link flaw may look small beside the risk involved. Public credit for the find pushes other researchers, in short, to test the same surfaces. OpenAI gains stronger defences, meanwhile, when researchers report flaws through proper channels.
Security teams at other firms should read this chain as a warning. A virtual keyboard sounds harmless at first glance. A maths library sounds far from account security, in the first place, yet both sat on the path to full page code execution. Teams that audit only main code can miss helpers that carry real risk.
ChatGPT users need no special action after the ChatGPT share link flaw patch, since OpenAI closed the hole centrally. People who shared links widely in recent weeks can review account activity for anything odd. Readers who follow browser flaws in developer tools will see a familiar pattern here. A small browser boundary failure opened a much larger door, by contrast, in that earlier coverage too.
Sandboxed previews now appear across chat apps, coding tools, and document editors. Each preview system promises isolation from untrusted code. Each vendor must still prove that promise with clear rules and fast fixes. The ChatGPT share link flaw story ends with a closed hole, a modest bounty, and a reminder that small components can carry large risk. Readers who track patched flaw research and core values for ChatGPT will find the same tension between open features and firm guardrails.
