Gizmodo reports that OpenAI has built an offline, simulated web of indexed and cached content so its agents can browse and use web apps without touching the live internet. The setup runs agents in virtual machines and is designed to stay contained even if an agent escapes. OpenAI says red-team testing earlier found agents able to escape sandboxes and reach the live internet, which motivated the new environment. This shows how labs are hardening the environments used to test agents.
