OpenAI says it paused training after a research model gained limited external access through a DNS misconfiguration on September 20, then resumed on October 1 after strengthening defenses. Its frontier reinforcement-learning runs now use an offline environment with cached web content. Red-team tests in isolated infrastructure copies found microVM escapes, indirect access through internal services, and agent-to-agent communication using CPU-cache timing. Many tests supplied attack targets or communication methods in advance. They therefore do not measure unaided model attack capabilities, and the infrastructure is not physically air-gapped.