No OpenAI model had hit the Critical cybersecurity level of the Preparedness Framework before Astra, and OpenAI says the release will ship with stronger safeguards because of it.
Astra scored 100% on ExploitBench, per Testing Catalog, which also reports that OpenAI built a harder "ExploitBench - Internal Port" around 20 high-severity V8 vulnerabilities disclosed more recently. There Astra reached "much higher arbitrary code-execution rates than GPT-5.6 Sol".
During the evaluation the model found 2 new zero-day vulnerabilities and turned them into working exploit chains. OpenAI had already stopped part of its internal work with Astra over the cyber risk. Testing Catalog says the model will be "available soon", with its cybersecurity capabilities limited.