Check Point Research found that the HTML preview for code blocks in ChatGPT left popup permissions enabled, letting content escape the sandbox and sidestep Lockdown Mode. In the Electron desktop app, popups worked without a user gesture, and the researchers showed a malicious GPT sending chat messages to an attacker's server. OpenAI says it removed the permissions in September and Check Point confirmed the method no longer works. Teams should not rely on Lockdown Mode alone and should check which AI services staff use with sensitive data.
