OpenAI's Cyber Risk for Astra Jumped From "High" to Potentially "Critical" in Eight Months

OpenAI flagged "high" cybersecurity risk for its upcoming models in December 2025. By August 7, 2026, it could not rule out "critical" capabilities for model Astra — autonomous zero-day discovery against hardened targets — and paused internal development. In this author's view, if this trajectory holds, the next checkpoint is whether such a model can be safely deployed under any monitoring regime.