OpenAI's Cyber Risk for Astra: From "High" to Possibly "Critical" in Eight Months

OpenAI flagged its upcoming models' cybersecurity risk as "high" in December 2025. By August 7, 2026, it could not rule out a "critical" level for Astra — meaning the model might autonomously find hidden flaws in well-protected computer systems — and halted development. In this author's view, the next question is whether any monitoring regime could safely allow its deployment.