What happened
OpenAI described its defense-in-depth approach to frontier AI risk across four layers — model alignment, security/sandboxing, safeguards, and monitoring — applied at each stage of the model lifecycle from highly sandboxed training environments to external Codex deployment, requiring a safety review to advance to the next stage. Following the July Hugging Face incident, it says it strengthened safeguards and alignment requirements, and now paces development, including pausing internal training and holding back models such as GPT-6.1 Astra where more time is needed to meet its safety standards, while relying on formal safety cases for frontier model training.
Why it matters
Shows how a frontier lab operationalizes risk gates, containment, and pausing in practice — a concrete reference point for regulators and enterprises designing their own model-risk gates and stage-gate controls.
Action needed
Compare the four-layer defense-in-depth model and its stage-gate safety reviews against your own AI model-risk governance framework.