OpenAI expands frontier-model monitoring to some external deployments

OpenAI has expanded monitoring for model misalignment beyond training, evaluation and internal use to some external deployments, DotAI Times reports. Coverage includes all reinforcement learning runs and evaluations for GPT-6 Astra and above, the software supporting their tool use, and all internal Codex traffic regardless of model. The system combines activation probes, model-based reviewers, investigation agents and automated investigations. High-priority alerts typically reach security and training teams within 15 minutes. Separately, OpenAI announced its Private Intelligence framework for enterprise data protection and agent permissions. Zero Data Retention applies to eligible API workflows, while Private Inference, which protects data during processing, is in preview.