Key Takeaways
OpenAI paused reinforcement learning training on its latest deployment models for two weeks and says its largest planned frontier run is still on hold.
The trigger was preliminary evidence that the upcoming Astra model may meet the Critical cybersecurity threshold in OpenAI's Preparedness Framework, a determination reached on August 7.
New monitoring inspects model activity at every sampled token, aims to alert within 30 minutes, and adds roughly 20% to the inference compute being watched.
OpenAI published an account this week of how it has slowed its own model development, and the detail underneath the headline is what makes it worth reading. The company temporarily paused reinforcement learning training on its latest models intended for deployment for two weeks while it hardened and red-teamed its research environments. The biggest run has not restarted. OpenAI wrote that its largest planned frontier run "remains on hold while we conduct smaller-scale training and evaluations".
Two things pushed it there. One was the OpenAI and Hugging Face security incident. The other was preliminary evidence that Astra, an upcoming model, may meet the Critical cybersecurity capability threshold under the company's Preparedness Framework. OpenAI made that call on August 7 and extended monitoring to every Astra inference run involving tools, not just training.
The monitoring itself is the part most coverage skipped. It starts with classifiers that inspect the model's internal activity at every sampled token and escalates to automated investigators looking for unauthorized access, data theft and attempts to defeat safeguards. OpenAI aims to raise an alert within 30 minutes. If safety, security and research teams cannot rule out a false positive inside another 30 minutes, they are expected to pause the activity. The company puts the overhead at about 20% of the inference compute being monitored, and says the security work has cost it real money and real delay.
That number is worth sitting with. A fifth of the compute now goes to watching, not building. WYDE covered Astra's arrival three weeks ago, and the federal framework it was headed toward before that. Help Net Security has the engineering breakdown, and Axios first reported the pause.
People Also Ask
Did OpenAI stop training its models? No. OpenAI paused reinforcement learning on its latest deployment models for two weeks and is keeping its largest planned frontier run on hold while smaller-scale training and evaluations continue.
What is the Critical cybersecurity threshold? Under OpenAI's Preparedness Framework, a model reaches it if it can independently find and build working zero-day exploits against hardened real systems, or run an end-to-end attack from a high-level goal.
What is Astra? Astra is an upcoming OpenAI model family. In August the company determined it may have critical cyber capability, which triggered the strictest security controls on any workload involving it.
How much does the new monitoring cost? OpenAI estimates roughly 20% of the inference compute being monitored, though it says the figure varies substantially across training and evaluation workloads.