Check Point bypassed ChatGPT's Lockdown Mode to silently send chat history to an outside server

Check Point Research · Company report

Read the source (opens in a new tab)
What happened
Check Point Research disclosed a flaw in ChatGPT's HTML preview sandbox that defeated Lockdown Mode, OpenAI's optional setting meant to stop prompt injections from exfiltrating user data.
Key points
  • Attackers could silently exfiltrate entire chat histories, phish for credentials and deliver malware, through a malicious prompt, a malicious GPT or a shared chat.
  • In the ChatGPT Classic desktop app the attack needed "zero user interaction"; a demo shows a malicious GPT leaking chat messages during an ordinary conversation.
  • Check Point says OpenAI told it on September 18, 2026 that it had deployed a fix by removing popup permissions from the sandbox; retesting confirmed the technique no longer works.
  • Check Point says built-in security features of AI apps "still offer no absolute guarantee of protection" as assistants run code and handle sensitive data.
Evidence
Security company research report with demonstration videos, published after OpenAI's fix. It does not report any use of the flaw by attackers.

See it in the latest record