Check Point Finds ChatGPT Flaw That Could Expose Chat Data

Cybersecurity researchers at Check Point Research uncovered a flaw in ChatGPT’s HTML page preview feature that allowed the Lockdown Mode protections to be bypassed. The issue centred on how the platform renders HTML content within code blocks inside nested iframes.

While the previews are sandboxed to prevent external network calls and unauthorized popups, the researchers showed that permissive sandbox tokens could enable popups to escape the sandbox via an auxiliary nested iframe, allowing rogue windows to load attacker-controlled content and potentially exfiltrate data.

In the desktop Electron-based client, exploiting the flaw could occur with zero user interaction, with an attacker’s script able to access complete chat logs and manipulate window titles, icons, and interfaces.

The impact, as described by Check Point, included the potential delivery of credential harvesting prompts and fake update notices designed to harvest credentials or deploy malware. OpenAI was informed of the vulnerability and subsequently addressed the weakness across ChatGPT distributions in September 2026, with the PoC exploits then failing in testing. The company advised enabling Lockdown Mode as a defence, noting that it limits automated internet retrieval and untrusted data processing.

The findings emphasise the continuing need for multi-layered isolation policies as AI platforms evolve, and demonstrate how seemingly small sandbox configuration changes can create pathways for data leakage and credential theft.