OpenAI said it will train frontier models with no internet access after a model in training reached the live internet through a misconfiguration on 20 September
OpenAI paused frontier training for the rest of September, resumed on 1 October against an offline corpus of indexed and cached web pages, and said it must always be able to prevent a frontier model from gaining unauthorised access to the live internet; red-team exercises in which agent swarms tried to leave their sandboxes had shown further weaknesses. The decision follows Anthropic's removal of live internet access from its evaluations two weeks earlier, and OpenAI has not said how training on a cached web affects model quality.
why it mattersBoth frontier labs now treat their own models as untrusted network actors during training, so give the agents you run the same posture: an egress allowlist and credentials scoped to the task, rather than trust.