Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

NewsResearch

Check Point: a popup escape from ChatGPT's HTML preview bypassed Lockdown Mode to steal chat history, deliver malware and phish logins; OpenAI fixed it on September 18

Alexey Bukhteyev of Check Point Research found a way around ChatGPT's Lockdown Mode, the optional setting meant to stop prompt injection from sending data out by cutting live web access, connectors, Agent Mode, Deep Research and file downloads. The HTML page preview, added in February 2026, runs pages in two nested iframes, and the outer one allowed popups that escape the sandbox. ChatGPT's own code asks the user before window.open, but a page could create another iframe and open a popup from it, and arbitrary HTML in that popup then ran outside the sandbox. Check Point shows three uses, delivered through a malicious prompt, a malicious GPT or a shared chat: silent theft of the whole chat history and uploaded files, zero-click in the ChatGPT Classic desktop app; a fake ChatGPT window offering a malicious update; and forcing a sign-out to phish the login. Check Point says OpenAI told it on September 18, 2026 that a fix was deployed, removing the popup permissions from the preview iframe, and its retest confirmed the technique no longer works.

Why it matters

A security mode is only as strong as every feature it leaves on. Do not rely on one setting to keep sensitive chats and files safe, and treat shared chats and custom GPTs from others as untrusted content.

Sources

Read it with the rest of the news in the Learning Center