BlockBeats report, August 8: OpenAI released a new blog post stating that internal tests show its upcoming Astra model has made significant progress in agent coding and cybersecurity. Based on its Preparedness Framework assessment, OpenAI stated it cannot rule out that the model has reached a 'critical' threshold of cyber capabilities.
Additionally, OpenAI explicitly clarified in the article that the Astra model was not involved in the recent Hugging Face security incident and continues benchmarking under the Preparedness Framework to manage frontier web capabilities. Accordingly, OpenAI has paused internal activities related to Astra that have not yet met these enhanced security controls.
The definition of "critical" capability is: the model can discover and exploit all severity-level functional zero-day vulnerabilities in a large number of real-world hardened critical systems without human intervention; or, based solely on high-level objectives, it can independently design and execute end-to-end novel cyberattack strategies against hardened targets. Previous models, such as GPT-5.6-Sol, were only evaluated as "high" level.