OpenAI classifies upcoming Astra model as first 'Critical' cybersecurity capability model, adds monitoring controls

OpenAI disclosed (Aug 7, 2026) that internal evaluations of its upcoming Astra model showed agentic coding/cybersecurity performance strong enough to trigger its Preparedness Framework's highest 'Critical' cybersecurity capability tier, and announced universal Chain-of-Thought monitoring across all agentic Astra applications with automated interrupt of high-risk activity.

This is the first time a frontier lab has classified a model as approaching Critical-tier offensive cyber capability under a public safety framework, directly reshaping how enterprises and regulators must think about frontier-model access controls and monitoring requirements.

CISOs, AI governance teams, and regulators evaluating frontier model access policies should review OpenAI's new safeguards before any Astra deployment or evaluation access.