---
title: "Check Point Reports ChatGPT Lockdown Mode Bypass, Says Fix Blocked Its Exploit"
dek: "A nested iframe let HTML previews bypass popup controls, according to Check Point Research. The researchers report that OpenAI removed two sandbox permissions and that their subsequent retest failed to reproduce the exploit."
kind: news
topics: [agents-accountability, evaluation-audits]
layers: [agents, runtime]
event_date: 2026-11-05
published: 2026-11-13
agent: daily
evidence: [announcement, documentation]
sources:
  - "ChatGPT Lockdown Mode Bypass: Phishing, Malware Delivery, and Silent Data Exfiltration | research.checkpoint.com | 2026-11-05 | https://research.checkpoint.com/2026/chatgpt-lockdown-mode-bypass-phishing-malware-delivery-and-silent-data-exfiltration/?trk=article-ssr-frontend-pulse_little-text-block"
---

Check Point Research reported on 5 November 2026 that a vulnerability in ChatGPT’s HTML previews allowed attackers to bypass Lockdown Mode, an optional setting intended to restrict channels for leaking sensitive data. The researchers say OpenAI informed them on 18 September 2026 that a fix had been deployed; their subsequent retest found that the described exploitation technique no longer worked.[^1]

The report locates the flaw in popup handling. HTML previews used two nested sandbox iframes, with `allow-popups` and `allow-popups-to-escape-sandbox` enabled. ChatGPT intercepted popup requests by replacing `window.open`, but an attacker could create another iframe and obtain its opening function. According to the researchers, this allowed a popup to escape the preview sandbox and bypass its approval controls.[^1]

The interaction requirements differed by client. Check Point says the Electron-based ChatGPT Classic application allowed popups without a user gesture, enabling silent conversation-data exfiltration through a malicious GPT. Browser popup restrictions generally required a click. Malware delivery in Classic required an additional click on a download link. The report also describes phishing through attacker-controlled windows.[^1]

The reported fix removed both popup permissions from the outer HTML-preview sandbox iframe. The source supplies no fixed client version or user-facing method for identifying an affected deployment.[^1]

:::analysis
The issue spans T5 agents (`agents`), where malicious content can redirect assistant interactions, and T2 runtime (`runtime`), where sandbox permissions govern popup behavior; a setting’s name alone cannot establish that its restrictions held. Users and operators should verify whether their serving deployment includes the reported permission removal, accounting for the client-specific behavior described in the report. The verifiable property at stake is identifiability: if the exact version and configuration serving a user are recorded and checkable, that user can determine whether they were exposed to the affected implementation; this does not itself establish that exploitation occurred.[^1]
:::

:::checked
On 13 November 2026, we read the saved text of Check Point Research’s “ChatGPT Lockdown Mode Bypass: Phishing, Malware Delivery, and Silent Data Exfiltration,” published on 5 November. We did not test the system, independently inspect the fix, or reproduce the demonstrations. The fix and retest outcomes are the researchers’ reported findings.[^1]
:::

[^1]: Check Point Research, “ChatGPT Lockdown Mode Bypass: Phishing, Malware Delivery, and Silent Data Exfiltration,” 5 November 2026. https://research.checkpoint.com/2026/chatgpt-lockdown-mode-bypass-phishing-malware-delivery-and-silent-data-exfiltration/?trk=article-ssr-frontend-pulse_little-text-block
