Skip to content- Sources: primary, discussion
- Summary: ChatGPT renders HTML previews of code blocks inside nested sandboxed frames. Check Point found that a further nested frame could reach an unmodified popup function, which opened content outside the sandbox and outside Lockdown Mode's limits on outbound requests. The flaw allowed chat data exfiltration, including zero-click exfiltration in the desktop app, through a crafted prompt, a malicious GPT or a shared chat. OpenAI fixed it on 2026-09-18.
- Why it matters: Lockdown Mode protected against data leaving through outbound requests, but this flaw in the HTML preview feature, outside the model, got around that boundary.