OpenAI pauses frontier RL training to tighten security after Hugging Face breach

OpenAI paused reinforcement learning training for two weeks and halted its largest planned frontier RL run after internal evaluations found its upcoming Astra model may meet 'critical' cybersecurity thresholds. New monitoring includes activation classifiers on every token, 30-minute alert response, and sandboxing, consuming ~20% of inference compute.

Featured · Amelia Glaese