OpenAI pauses frontier training after Astra nears Critical cyber threshold

OpenAI paused reinforcement-learning training for two weeks and held back its largest planned frontier run after preliminary internal evidence suggested Astra could reach the Critical cybersecurity tier of its Preparedness Framework — capability that could uplift a sophisticated attacker without custom tooling. Monitoring now covers all tool-enabled training and evaluations at GPT-5.6 Sol capability or above, plus all Astra inference, targeting alerts within 30 minutes. A lab delaying its own frontier run on cyber risk is the first concrete case of a capability threshold, rather than a release calendar, setting the schedule.

Source: openai.com ↗