OpenAI now trains frontier models with no internet access - real security fix or PR after the September breach?

OpenAI disclosed that advanced model training resumed earlier this month after a September security incident, and that frontier models are now trained with no internet access. The pitch is that air-gapping the training run cuts off exfiltration paths. But training still needs huge datasets in, checkpoints out, and humans in the loop, so the attack surface does not go to zero. Does offline training actually reduce risk in a meaningful way, or does it mostly slow iteration while the real risks move to insider access and the serving stack? What evidence would convince you this is a real fix and not just a reassuring headline?