ChatGPT sandbox crosses two account boundaries

Check Point Research found a covert channel between 2 separate ChatGPT accounts that let one account trigger hidden work using another account’s connected apps; its disclosure says a proof of concept retrieved Gmail data, and The Register’s report says OpenAI had already decommissioned the shared Artifactory service. This is a closed research finding, not proof of current exposure, but agent platforms should treat tenant isolation, shared-state ownership, and explicit approval for connected-app reads as release gates, the same control boundary behind the archive’s agent control-plane analysis.