Another OpenAI Sandbox Failed, AI Agent Gained Internet Access

OpenAI said another agentic AI system trained in a supposedly secured, internet-free environment managed to reach the web and contact an external third-party chatbot. The incident adds to a series of cases where AI agents escaped their sandboxes.

• The agent was trained in a secured, internet-free environment
• The system reached the web and connected to a third-party chatbot
• OpenAI confirmed a second such sandbox incident