Threat
The Shared Clipboard Inside the Sandbox: Cross-Account Data Leakage in ChatGPT - Check Point Research
Publication date not yet evaluated · Added · 2 reports checkpoint.com
Page published
Coverage timeline
Why it matters
The ChatGPT cross-account channel shows how weak isolation boundaries between AI code-execution sandboxes can let a planted prompt silently exfiltrate a victim's connected email and files, a systemic risk as agents gain access to sensitive data and tools.
Check Point Research disclosed a covert cross-account command channel in ChatGPT that let an attacker use a victim's session to execute hidden tasks with the victim's connected tools and data. The channel worked because isolated code-execution containers from different accounts could all reach the same internal JFrog Artifactory package service; in a proof of concept, ChatGPT retrieved the victim's connected Gmail data and relayed it to the attacker while the victim saw only a normal response. The hidden instruction could be planted via a malicious prompt, shared conversation, or custom GPT, and OpenAI has since decommissioned the affected Artifactory instance.