The new safeguards include stronger network isolation, more detailed monitoring during model development, and greater emphasis on alignment and security during post-training. The company said that under the revised architecture, no individual breach of a workload or supporting service would on its own be sufficient to open a path to the internet or other internal networks. The computational overhead of the monitoring system is pegged at around 20% of the workload it covers, and the company has set a target of generating alerts no more than 30 minutes after suspicious activity is detected, according to TechCrunch.