OpenAI details staged risk management across a frontier model's lifecycle
What happened
On November 18, OpenAI published an explainer on how it manages risk for frontier AI models in stages, from training to public deployment. Models are barred from live internet access during research. In internal and external deployment, alignment evaluations, system safeguards and monitoring tighten step by step. Each stage needs a safety review before it proceeds, and OpenAI can pause training or hold back releases such as GPT-6.1 Astra when safety evidence is insufficient. The post sets out the requirements and the pause conditions but does not report any model actually paused or delayed.
Written by AI from the coverage · updated 2 days ago
Coverage
Follow the reports to see the story from different sides.
- OpenAI NewsPickHow OpenAI manages frontier AI risk from training to deployment
OpenAI laid out how it manages risk in stages as a frontier model moves from training to public deployment, pausing training or holding back releases such as GPT-6.1 Astra when safety evidence falls short. Research runs without live internet access, and internal and external deployment add alignment evaluations, system safeguards and monitoring. Each stage needs a safety review before it proceeds.
Heat over time
Heat now 3·Comparable peak 6(Nov 19 06:00)·Comparable change over 24 hours -49%
The trend only compares accounts observed without gaps, so its range may be smaller than the current heat. Hover or tap the chart for each hour; the left and right arrow keys step through it.