ChatGPT Lockdown Mode Bypass Exposed, Fixed
Check Point Research disclosed a Lockdown Mode bypass using ChatGPT’s HTML preview popup handling to enable silent chat-data exfiltration, phishing, and malware delivery. OpenAI removed the enabling sandbox permissions on September 18, and researchers confirmed the technique no longer worked.
This is a reminder that AI security depends on the entire application stack, not just model guardrails. Lockdown Mode reduced outbound paths, but an overlooked UI sandbox behavior created another route around them.
- –The exploit abused nested HTML-preview frames and popup handling to escape sandbox restrictions.
- –In the desktop app, malicious content could silently exfiltrate chat messages without a user gesture.
- –The same trusted interface could host convincing fake update prompts or credential-phishing pages.
- –Developers should threat-model runtimes, sandboxes, connectors, and rendering components alongside prompt injection defenses.
- –OpenAI’s own documentation now frames Lockdown Mode as risk reduction, not a guarantee against data exfiltration. [OpenAI](https://help.openai.com/en/articles/20001061-lockdown-mode)
DISCOVERED
1d ago
2026-11-06
PUBLISHED
1d ago
2026-11-06
RELEVANCE
AUTHOR
IntCyberDigest
