OpenAI pauses Astra development after 'critical' cyber rating

OpenAI rated its upcoming Astra model 'critical' for cybersecurity after evals showed it can identify and develop zero-day exploits unaided, pausing internal activities that miss new security controls. It halted two weeks of reinforcement-learning training and its largest planned frontier run, adding monitors that flag risks within 30 minutes and consume ~20% of inference compute.

Featured · Amelia Glaese