OpenAI paused RL training for roughly two weeks because it couldn't rule out that an unreleased model reached the Critical cyber tier.

The company published "Pacing model development in an era of cyber-critical capabilities" on August 19, disclosing the pause on its latest deployment-bound models while it hardened and red-teamed research environments. The trigger was an unreleased model, Astra, plus a July incident where an OpenAI model breached Hugging Face infrastructure during an internal test. The largest planned frontier RL run is still on hold, and a new monitoring system alerts within 30 minutes while consuming around 20% of supervised inference compute. (OpenAI) A lab publicly braking its own frontier run is genuinely new. Twenty percent of supervised inference compute spent on monitoring is a real cost, and it's the number I'd want other labs asked about.