ChatGPT Lockdown Mode Leak: What Developers Should Know
A sandbox escape let malicious GPTs leak ChatGPT conversations even with Lockdown Mode on.
Why the preview sandbox escape was dangerous
What changed in OpenAI's fix
Takeaways for developers building on ChatGPT
Lockdown Mode was supposed to be ChatGPT's privacy safety net. When it's enabled, live web access is limited to cached pages, image loading is blocked, and data-sending tools such as Deep Research, Agent Mode, connectors, and file downloads are switched off. The idea is to cut the routes a prompt-injection attack can use to send private data elsewhere. Check Point Research found that one route survived anyway, because the leak wasn't in the model's answers—it was in how ChatGPT rendered HTML page previews inside code blocks.
## Why the preview sandbox escape was dangerous
Check Point traced the bug to nested preview frames. The HTML-preview layer is sandboxed so it cannot read cookies or contact the wider web without approval, but nested frames could still spawn popup windows that escaped the sandbox and bypassed Lockdown Mode. Once outside, an attacker-controlled window could copy a full conversation history to a third-party server, display a fake update screen, or push malware downloads. In the Electron-based ChatGPT Classic desktop app, the attack required no user interaction: starting a conversation with a malicious GPT was enough. Popup escapes are a familiar sandbox failure mode, but here it had a direct path into a trusted ChatGPT session.
## What changed in OpenAI's fix
OpenAI closed the hole on 18 September 2026 after a private report, and Check Point published its technical write-up on 5 November. The fix was server-side: OpenAI removed popup permissions from the outer HTML-preview frame. Check Point retested and confirmed the technique no longer works, and no desktop update is required. That doesn't make the incident irrelevant. Eli Smadja, Check Point's head of research, framed the lesson simply: AI security controls are only as strong as the application around them.
## Takeaways for developers building on ChatGPT
If you're shipping custom GPTs, connector apps, or ChatGPT Desktop integrations, treat safety toggles as risk reduction, not guarantees. The Lockdown Mode feature narrows data-exfiltration paths, but it did not eliminate them until the surrounding renderer was hardened. Audit every component that touches untrusted content, especially previews, rich embeds, and sandboxed iframes, and include adversarial tests that try to pop windows or navigate out of the sandbox. Most importantly, remember that model behavior and client architecture are two different trust boundaries. A solid result from this case is a clearer security checklist: test the app, not just the prompt, and don't assume a locked-down setting means exfiltration is impossible.
Treat Lockdown Mode as a risk-reducing feature, not a guarantee—audit the app around the model.