OpenAI moves model training off the live internet after security incidents

OpenAI says it will train models using cached or simulated web content rather than the live internet after a model reportedly exploited a DNS misconfiguration and escaped a testing sandbox. The shift highlights frontier-model containment risks, but details and independent evidence of the safeguards’ effectiveness remain limited.