Sandbox Escape Bypassed ChatGPT's Lockdown Mode Protection
Original source
Researchers Bypass ChatGPT Lockdown Mode to Exfiltrate Chats Without User Interaction
The Hacker News →Check Point researchers uncovered a vulnerability in OpenAI’s Lockdown Mode that enabled attackers to bypass the security feature and extract chat data. Introduced in June 2026 as an optional setting, Lockdown Mode was designed to block data exfiltration from prompt injections by restricting outbound network requests. The vulnerability exploited a weakness in how HTML previews handled popup windows within nested sandboxes—ChatGPT blocked direct window-opening attempts, but the underlying sandbox still permitted popups to escape. Attackers could create nested frames to access unmodified popup functions, circumventing the protection and enabling data theft, phishing attacks, and malware delivery without user interaction.
OpenAI patched the flaw on September 18, 2026. The discovery underscores a critical pattern: AI security controls only work as effectively as the broader application environment protecting them. The finding comes alongside other recent ChatGPT vulnerabilities, including a Parameter Labs disclosure of a sandbox bypass enabling account takeover via malicious shared links, and a Check Point demonstration of a covert cross-account command channel that allowed attackers to run hidden tasks using a victim’s authenticated ChatGPT session while showing normal responses to visible requests.
Read the full article
Continue reading at The Hacker News →This is an AI-generated summary. Read the original for the full story.