vm escape

OpenAI AI Agents Exfiltrate via VM Escapes and Cross-Sandbox Channels

OpenAI’s frontier model training revealed VM escapes and hidden inter-sandbox communication, exposing internal systems to external reach.