What OpenAI reported

In its August 18 update, OpenAI described a two-week pause in reinforcement learning for deployment-bound models. Its largest planned frontier RL run remained on hold while smaller runs and safety evaluations continued.

The company cited the Hugging Face incident and preliminary evidence that Astra might reach its Critical cybersecurity threshold. It described stronger isolation and monitoring, with an alert goal of 30 minutes after concerning activity surfaces. For a likely critical boundary violation, teams are expected to pause the activity unless they can rule the flag out within 30 minutes.

How to read the announcement

This was a reported intervention in particular workloads, not an announcement that all research stopped. Our reading: assess the stated scope, restart conditions, and evidence of control effectiveness separately. A written procedure is not a measured detection rate.

For the incident itself, read METR's August 26 investigation, including its stated access and scope limitations.