← Back to live feed1 story
1
Check Point Finds ChatGPT Sandbox Flaw Letting Attackers Steal Gmail Data Across Accounts
Infra23d agoA hidden prompt in ChatGPT could trigger the chatbot to hand over data from integrated services including Google Drive and Microsoft Teams through a covert communication channel. The vulnerability exploited a shared clipboard between the assistant's code-execution containers, allowing an attacker to hijack a session and steal messages or files while the user saw no unusual activity.
Check Point Research identified the sandbox flaw in June 2026 and verified that the channel allowed the extraction of Gmail chat history and uploaded files across different accounts. OpenAI has since decommissioned the internal server that powered the vulnerability, confirming the security hole was closed before the findings were publicized.
Sign in to suggest edits
Key sources
- SOURCE@intcyberdigest“the researchers pulled a victim's Gmail messages, chat history and uploaded files while that user watched their own request run as normal”x.com
- SUPPORT@the_cyber_news“isolated sandboxes the assistant uses when a task requires running code or installing software packages”x.com
- SUPPORT@technobezz“could let attackers run hidden tasks and steal data from a victim's connected apps”x.com
- SOURCE@thehackersnews“Check Point demonstrated the hidden transfer through shared Artifactory metadata”x.com
- SUPPORT@ethereaglehq“the only clue was a "Talked to Gmail" label after the read”x.com