OpenAI states that based on alignment research, monitoring, and internet infrastructure security, frontier model training should run with no live internet access until safeguards justify broader access. After a September 20 DNS misconfiguration let an internal research model reach limited live internet during training, OpenAI paused frontier RL training and evaluation, hardened infrastructure, and resumed frontier RL and internal inference on October 1. The post describes sandboxes, microVMs, monitoring, and an offline world of indexed/cached web content for RL runs. Red-teaming with agent swarms found vulnerabilities that were patched; agents also demonstrated cache-timing side-channel communication without network connections.